⚡Megatron LLM Hub
All digests
Website

rss-website

Top3

未认证即可利用,暴露 27918 端口即受影响,7.5.0 需尽快升级到 7.5.1。

PSIRT Arbitrary file write in grpc phQueryMaster Summary An Improper limitation of a pathname to a restricted directory ('path traversal') [CWE-22] in FortiSIEM may allow an unauthenticated attacker to execute unauthorized code or commands via specially crafted packets. Version Affected Solution FortiSIEM 7.6 Not affected Not Applicable FortiSIEM 7.5 7.5.0 Upgrade to 7.5.1 or above Workaround : - Restricting access on tcp port 27918 prevents the issue. FortiSIEM Cloud is NOT affected by this vulnerability. Acknowledgement Internally discovered and reported by Gwendal Guégniaud of Fortinet Product Security team. Virtual Patch No Timeline 2026-10-13: Initial publication IR Number FG-IR-26-180 Published Date Oct 13, 2026 Component OTHERS Severity Critical Discovered Internal Attack Type Unauthenticated Known Exploited No CVSSv3 Score 9.4 Impact Execute unauthorized code or commands CVE ID CVE-2026-106173 Download CVRF CSAF

rce路径穿越固件cve @forti-ir source ↗

未认证即可绕过,5.0.0-5.0.6 与 4.4.0-4.4.9 均受影响,已有修复版本。

PSIRT Improper access control and session management in the web management interface Summary An Authentication Bypass Using an Alternate Path or Channel vulnerability [CWE-288] in FortiSandbox GUI may allow a remote unauthenticated attacker to bypass authentication via crafted HTTP requests. Version Affected Solution FortiSandbox 5.2 Not affected Not Applicable FortiSandbox 5.0 5.0.0 through 5.0.6 Upgrade to 5.0.7 or above FortiSandbox 4.4 4.4.0 through 4.4.9 Upgrade to 4.4.10 or above FortiSandbox PaaS 5.2 Not affected Not Applicable FortiSandbox PaaS 5.0 5.0.0 through 5.0.6 Upgrade to 5.0.7 or above FortiSandbox PaaS 4.4 4.4 all versions Migrate to a fixed release Acknowledgement Internally discovered and reported by Adham El Karn of Fortinet Product Security team. Fortinet is also pleased to thank Jason McFadyen of TrendAI Research for reporting this vulnerability under responsible disclosure. Virtual Patch No Timeline 2026-10-13: Initial publication IR Number FG-IR-26-185 Published Date Oct 13, 2026 Component GUI Severity Critical Discovered Internal Attack Type Unauthenticated Known Exploited No CVSSv3 Score 9.6 Impact Execute unauthorized code or commands CVE ID CVE-2026-107656 Download CVRF CSAF

认证绕过cve绕过固件 @forti-ir source ↗

集中管理平台一旦被低权限账号攻陷即可任意执行命令,7.4/7.6 均在受影响范围。

PSIRT Multiple Arbitrary File Read/Write via SQL Injection Summary An Improper Neutralization of Special Elements used in an SQL Command ('SQL Injection') vulnerability [CWE-89] in FortiAnalyzer, FortiManager, FortiManager Cloud and FortiAnalyzer Cloud may allow a low privileged user to execute arbitrary commands via crafted requests. Version Affected Solution FortiAnalyzer 8.0 Not affected Not Applicable FortiAnalyzer 7.6 7.6.0 through 7.6.4 Upgrade to 7.6.5 or above FortiAnalyzer 7.4 7.4.0 through 7.4.7 Upgrade to 7.4.8 or above FortiAnalyzer 7.2 7.2 all versions Migrate to a fixed release FortiAnalyzer Cloud 7.6 7.6.2 through 7.6.4 Upgrade to 7.6.5 or above FortiAnalyzer Cloud 7.4 7.4.1 through 7.4.7 Upgrade to 7.4.8 or above FortiAnalyzer Cloud 7.2 7.2.1 through 7.2.11 Upgrade to 7.2.12 or above FortiManager 8.0 Not affected Not Applicable FortiManager 7.6 7.6.0 through 7.6.4 Upgrade to 7.6.5 or above FortiManager 7.4 7.4.0 through 7.4.6 Upgrade to 7.4.7 or above FortiManager 7.2 7.2 all versions Migrate to a fixed release FortiManager Cloud 8.0 Not affected Not Applicable FortiManager Cloud 7.6 7.6.2 through 7.6.4 Upgrade to 7.6.5 or above FortiManager Cloud 7.4 Not affected Not Applicable FortiManager Cloud 7.2 Not affected Not Applicable Acknowledgement Internally reported and discovered by Jaguar Perlas of Burnaby Infosec team and David Maciejack from Fortinet Product Security team. Virtual Patch No Timeline 2026-10-13: Initial publication IR Number FG-IR-26-178 Published Date Oct 13, 2026 Component GUI Severity High Discovered Internal Attack Type Authenticated Known Exploited No CVSSv3 Score 7.9 Impact Execute unauthorized code or commands CVE ID CVE-2025-46773 CVE-2026-23705 CVE-2026-23706 CVE-2026-23707 Download CVRF CSAF

sqlircecve提权 @forti-ir source ↗
Must-see4

加载不可信 checkpoint 即触发,0day 无补丁,仿真资产成为新攻击面

(0Day) Genesis World load_checkpoint Deserialization of Untrusted Data Remote Code Execution Vulnerability October 13th, 2026 ZDI-26-756 ZDI-CAN-32128 CVE ID CVSS Score 7.8 AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H Affected Vendors Genesis World Affected Products Genesis World Vulnerability Details This vulnerability allows remote attackers to execute arbitrary code on affected installations of Genesis World. Interaction with this library is required to exploit this vulnerability but attack vectors may vary depending on the implementation. The specific flaw exists within the processing of checkpoint files. The issue results from the lack of proper validation of user-supplied data, which can result in deserialization of untrusted data. An attacker can leverage this vulnerability to execute code in the context of the current process. Additional Details 07/16/26 - ZDI emailed a Genesis World contributor to request a secure channel for vulnerability disclosures 07/28/26 - ZDI asked the vendor to update their security policy via a GitHub issue 07/28/26 - the vendor closed the issue 08/06/26 - ZDI followed up requesting the vendor’s PSIRT contacts 09/11/26 - ZDI notified the vendor of the intention to publish the case as a 0-day advisory -- Mitigation: Given the nature of the vulnerability, the only salient mitigation strategy is to restrict interaction with the product Disclosure Timeline 2026-09-25 - Vulnerability reported to vendor 2026-10-13 - Coordinated public release of advisory 2026-10-13 - Advisory Updated Credit Aaron Luo of VicOne LAB R7 Back to Advisories

rce反序列化0daycve @zdi-advisories source ↗

ICS 是近年被在野利用最多的边界设备之一,需确认受影响版本与补丁状态。

Security is foundational to how Ivanti builds, tests, and supports our products. We continue to strengthen our secure product development practices through vulnerability discovery, responsible disclosure processes, security testing, threat-informed hardening, and coordinated response programs. Additionally, we recognize the importance of clear, timely communication when security issues arise. To that end, today we are disclosing a vulnerability in limited versions of Ivanti Connect Secure. At the time of disclosure, Ivanti is aware of limited exploitation in the wild. We are working directly with those organizations while making information, updates and remediation available to the broader customer community. If you are running one of the impacted versions, Ivanti recommends you act immediately. Additional information related to the impacted versions and remediations can be found in the Security Advisory: Ivanti Connect Secure Customers should review the security advisory, upgrade to the fixed version as soon as possible, and run the Integrity Checker Tool regularly. As with any security update, applying the fix is an essential step; however, it should be part of a layered security approach that includes strong access controls, network monitoring, log review, and ongoing validation of appliance integrity. Patch Tuesday Disclosures As part of our regular security program, Ivanti releases standard security patches on the second Tuesday of every month. At the time of disclosure, Ivanti is not aware of exploitation in the wild, and the fixes have been released. More information can be found in the Security Advisories: Ivanti Neurons Patch for Intune Xtraction Our Support team is always available to help customers and partners should they have any questions. Cases can be logged via the Ivanti Innovators Hub (login credentials required). Want to stay up to date on Ivanti Security Advisories? Paste https://www.ivanti.com/blog/topics/security-advisory/rss into your preferred RSS reader / functionality in your email program.

vpn补丁边缘设备事件 @ivanti-advisory source ↗

硬编码密钥无法靠配置缓解,5.0–5.2 多个版本受影响须升级。

PSIRT Fabric API Authentication bypass Summary A Use of Hard-coded Cryptographic Key [CWE-321] in FortiSandbox and FortiSandbox Cloud API may allow an unauthenticated attacker to access some endpoints via specially crafted HTTP requests Version Affected Solution FortiSandbox 5.2 5.2.0 through 5.2.1 Upgrade to upcoming 5.2.2 or above FortiSandbox 5.0 5.0.0 through 5.0.6 Upgrade to 5.0.7 or above FortiSandbox 4.4 4.4 all versions Migrate to a fixed release FortiSandbox Cloud 5.2 5.2.0 through 5.2.1 Upgrade to upcoming 5.2.2 or above FortiSandbox Cloud 5.0 5.0.0 through 5.0.6 Upgrade to upcoming 5.0.7 or above FortiSandbox Cloud 4.4 4.4 all versions Migrate to a fixed release Acknowledgement Fortinet is pleased to thankSina Kheirkhah (SinSinology) of watchTowr (watchTowrCyber) for reporting this vulnerability under responsible disclosure. Virtual Patch No Timeline 2026-10-13: Initial publication IR Number FG-IR-26-179 Published Date Oct 13, 2026 Component API Severity Medium Discovered External Attack Type Unauthenticated Known Exploited No CVSSv3 Score 4.9 Impact Escalation of privilege CVE ID CVE-2026-106172 Download CVRF CSAF

授权绕过硬编码密钥固件 @forti-ir source ↗

暴露在公网的 FortiGate 与 SSL VPN 凭据被大规模收集,泄露即边界失守。

Oct 7, 2026 FortiBleed: SafeBreach Coverage for Joint Cybersecurity Advisory JCSA-20261006-01 Learn how SafeBreach maps JCSA-20261006-01 to simulations that test your exposure to the FortiBleed credential-harvesting campaign against Fortinet FortiGate firewalls and SSL VPNs. Summary The FBI and US Secret Service issued Joint Cybersecurity Advisory JCSA-20261006-01 on October 6, 2026. It covers FortiBleed, a global operation that harvests credentials and sells access to internet-exposed Fortinet FortiGate firewalls and SSL VPN gateways. The activity was observed between June 18 and July 23, 2026, and the advisory warns that scanning is ongoing. SOCRadar has verified more than 86,644 FortiGate devices with working stolen credentials across 194 countries. The attackers use password spraying and credential stuffing to get in, steal FortiOS user databases and session tokens, and crack weak SHA-256 hashes offline on rented GPUs. They then create rogue admin accounts, map Active Directory and lock out the real administrators. The campaign doesn’t exploit a vulnerability. It relies on reused, leaked, and weakly stored credentials, and it has given INC/Lynx and Payload ransomware affiliates a way in. This post outlines the campaign’s TTPs mapped to MITRE ATT&CK. It also summarizes the advisory’s IOCs, SafeBreach simulations that validate whether your controls detect and block the campaign’s brute force, password spraying, Active Directory discovery, exfiltration, and command-and-control communication, and steps to harden FortiGate environments. Author: Noam Sagiv, Product Manager On October 6, 2026, the Federal Bureau of Investigation (FBI) and the US Secret Service (USSS) released a joint Cybersecurity Advisory titled “FortiBleed Operations Continue Targeting Exposed Systems Leading to Reports of Lockouts” (Product ID: JCSA-20261006-01) . This advisory provides critical intelligence on the FortiBleed campaign, including its tactics, techniques, and procedures (TTPs), as well as indicators of compromise (IOCs) observed in attacks against Fortinet FortiGate devices. The campaign is global and sector-agnostic: the advisory lists all 16 US critical infrastructure sectors as affected, from energy and healthcare to financial services, government facilities, and information technology. Organizations are experiencing lockouts when the actors disable or modify legitimate accounts on compromised devices. What is the FortiBleed Threat? FortiBleed is a credential-harvesting and initial-access-broker operation focused on the edge of the enterprise network. Rather than exploiting a software flaw, the actors abuse the credentials that protect FortiGate firewalls and SSL VPN gateways. These credentials are often reused, previously leaked, or stored using legacy SHA-256 hashing that is practical to crack at scale. The operation is industrialized. The actors scan the internet for exposed FortiGate SSL VPN portals, run password spraying and credential stuffing against them, and pull FortiOS user databases and session tokens from the devices they reach. Harvested hashes are fed into a distributed cracking pipeline built on Hashcat and Hashtopolis, running on rented GPU infrastructure. Scale is what sets FortiBleed apart. SOCRadar verified more than 86,644 FortiGate devices with working stolen credentials across 194 countries, and independent technical analysis from CloudSEK of an exposed FortiBleed open directory shows the tooling the actors left behind. Fortinet’s PSIRT has also published its own analysis of the reported credential compromise, noting that the same actor has reportedly used brute force credential harvesting against other vendors’ devices, so the exposure is not limited to FortiGate. The downstream impact extends well beyond the firewall. Once inside, the actors create new administrative accounts for persistence, enumerate Active Directory to prepare for lateral movement, and remove legitimate users’ access. According to the advisory, this access chain has served as an initial entry point for ransomware affiliates, including INC/Lynx and Payload ransomware. FortiBleed TTPs Mapped to MITRE ATT&CK The advisory maps FortiBleed activity to the following MITRE ATT&CK techniques. Resource Development Acquire Infrastructure ( T1583 ): The actors rented GPU infrastructure to crack password hashes at scale. Reconnaissance and Initial Access Active Scanning ( T1595 ): The actors scanned the internet for exposed FortiGate SSL VPN portals. Exploit Public-Facing Application ( T1190 ): The actors targeted publicly exposed FortiGate infrastructure. Password Spraying ( T1110.003 ): The actors attempted a small set of common passwords across many accounts. Credential Stuffing ( T1110.004 ): The actors used credentials obtained from prior breaches. Persistence Create Local Account ( T1136.001 ): The actors created new administrative accounts on FortiGate devices, often with names designed to blend in, such as variations of “admin,” “support,” and “fortinet.” Credential Access OS Credential Dumping ( T1003 ): The actors extracted FortiOS user databases and session tokens. Password Cracking ( T1110.002 ): The actors used Hashcat and Hashtopolis for distributed offline cracking of harvested hashes. Discovery Account Discovery ( T1087 ): The actors enumerated Active Directory to support lateral movement. Exfiltration Exfiltration Over C2 Channel ( T1041 ): Harvested credentials were transmitted to attacker-controlled servers. Impact Account Access Removal ( T1531 ): The actors locked, disabled, or altered account access, producing the lockouts the advisory warns about. The advisory does not reference any Common Vulnerabilities and Exposures (CVEs). The attack chain depends on credential exposure and weak password storage rather than a specific FortiOS vulnerability. FortiBleed IOCs The advisory provides IOCs, including: Attacker infrastructure IP addresses, including a command-and-control (C2) server, proxy nodes, a beacon relay, and a Hashtopolis cracking server. More than a dozen IP addresses observed conducting attacks between June 18 and July 23, 2026, with observation dates for each. Account names the actors created or used on compromised FortiGate devices (e.g., forticloud-sync, support_fortinet, adminsslvpn). For the full list of IOCs, refer to the advisory PDF . Organizations should audit FortiGate local and administrator accounts against the advisory’s account-name list and review firewall, VPN, and authentication logs for connections from the listed IP addresses. SafeBreach Coverage & Playbook Attack Updates Existing Behavioral Coverage SafeBreach already simulates several of the attacker behaviors detailed in this campaign. The following 11 existing attacks are now mapped to FortiBleed: Brute force: #173 Brute force attack over SSH protocol #11771 Brute force attack over SSH protocol against a live service #258 Brute force attack over the HTTP/S protocol Password spraying: #11666, #11667 and #11668 Password Spray on Fake Domain Accounts via SMB, LDAP and Kerberos Active Directory discovery: #11543 Extract domain users using net.exe (Windows) #6578 Discover domain users using LDAP method #8890 Dump Active Directory data using ldifde Exfiltration: #107 Covert data asset exfiltration using HTTP/S POST #110 Covert data asset exfiltration using HTTPS POST NOTE: Simulation #11771 is an Advanced Action because it repeatedly authenticates against a live service, so schedule it accordingly. These simulations are mapped to MITRE ATT&CK and can be used immediately to validate detection and response capabilities across your infrastructure. New IOC-Based Coverage To address the newly released IOC guidance, we have added the following attack: #11813 Communication with FortiBleed (FBI/USSS JCSA-20261006-01) using HTTP, which tests whether your controls detect and block HTTP communication with the IP addresses the advisory attributes to the campaign’s command-and-control and attack infrastructure. The indicator is carried inside a simulated request, and no traffic reaches the actors’ infrastructure. What You Should Do Now SafeBreach customers can now validate their security controls against these TTPs in multiple ways. Method 1 – SafeBreach Scenarios: Navigate to the SafeBreach Scenarios page and choose CYBERSECURITY ADVISORY FortiBleed. Method 2 – Attack Playbook: From the Attack Playbook, filter by the threat name CYBERSECURITY ADVISORY FortiBleed or search for the attack IDs listed above. Method 3 — Known Threats Series: Open the Known Threats Series report, select the CYBERSECURITY ADVISORY FortiBleed report, and run the simulations. Additional Advisory Steps How to Protect FortiGate Devices from FortiBleed The authoring agencies recommend the following: Reduce the management attack surface. Restrict external management of FortiGate devices via trusted hosts, a local-in policy, or, best of all, remove internet-facing administration altogether. Terminate sessions and reset credentials. End all active administrative and VPN sessions, then reset all Fortinet VPN and administrative passwords, prioritizing internet-facing systems. Require phishing-resistant MFA. On all remote access and administrative accounts, require phishing-resistant MFA. Move off legacy hashing. Confirm administrator passwords use the Password-Based Key Derivation Function 2 (PBKDF2) rather than legacy SHA-256. See Fortinet’s guidance . Validate configurations against a baseline. Review firewall and VPN users for unauthorized changes, and remove any unknown REST API keys. Review logs. Examine firewall, VPN, authentication, and domain controller logs for suspicious activity. How to Implement Proactive Threat Monitoring Audit FortiGate accounts. Flag any administrator account that does not match your baseline, especially names mimicking Fortinet support or cloud-sync functions. Alert on authentication patterns. Watch for many accounts failing against a small set of passwords, and for successful VPN logins from the advisory’s listed IP addresses. Watch for lockouts. Treat unexpected administrator lockouts or disabled accounts on edge devices as a possible compromise, not a help desk ticket. Monitor for directory enumeration. Investigate Active Directory account discovery originating from VPN address pools or firewall-adjacent hosts. Inventory REST API keys. Track creation of new FortiOS API keys and remove any that cannot be attributed. Stay Ahead with SafeBreach To see how your controls hold up against the FortiBleed campaign, sign into SafeBreach and run the 12 simulations mapped to JCSA-20261006-01. In addition to testing your defenses against brute force, password spraying, and Active Directory enumeration, you can go a step further with SafeBreach Propagate . Propagate enables you to assess how attackers could pivot across your environment post-compromise. It maps high-risk attack paths, visualizes lateral movement, and prioritizes remediation efforts based on exposure to your most critical assets. Learn more in the SafeBreach Propagate solution brief . About the Author Noam Sagiv is a Product Manager at SafeBreach leading attack content for the SafeBreach CTEM Platform. Before SafeBreach, she spent five years in Israel’s Unit 8200, where she led teams building data-driven intelligence products used by more than 10,000 users. Frequently Asked Questions What is FortiBleed? FortiBleed is a credential-harvesting campaign that targets internet-exposed Fortinet FortiGate firewalls and SSL VPN gateways. The FBI and US Secret Service described it in Joint Cybersecurity Advisory JCSA-20261006-01, released October 6, 2026. FortiBleed was observed between June 18 through July 23, 2026, and SOCRadar verified more than 86,644 FortiGate devices with working stolen credentials across 194 countries. The attackers don’t exploit a CVE. Instead, they use leaked and reused credentials, crack weak SHA-256 password hashes offline, create rogue administrator accounts, and lock out legitimate admins. Ransomware affiliates, including INC/Lynx and Payload, have used this access to get in. What is Joint Cybersecurity Advisory JCSA-20261006-01? JCSA-20261006-01 is a joint Cybersecurity Advisory released on October 6, 2026 by the FBI and the US Secret Service. The advisory describes FortiBleed, an active, global credential-harvesting campaign targeting internet-exposed Fortinet FortiGate firewalls and SSL VPN gateways. JCSA-20261006-01 warns that FortiBleed actors are locking legitimate administrators out of their devices and that the access has been used by ransomware affiliates. All 16 US critical infrastructure sectors are listed as affected. Which systems are affected, and is there a CVE? FortiBleed targets internet-exposed Fortinet FortiGate firewalls and FortiOS SSL VPN gateways. JCSA-20261006-01 cites no CVE. The FortiBleed campaign relies on reused and leaked credentials, password spraying, and offline cracking of FortiOS password hashes stored with legacy SHA-256. Organizations exposing FortiGate administration or SSL VPN portals to the internet, without phishing-resistant MFA or PBKDF2 password hashing, are at greatest risk. Who is behind FortiBleed, and what are they doing? FortiBleed is a credential-harvesting and access-broker operation. According to JCSA-20261006-01, the FortiBleed actors scan for exposed FortiGate portals, harvest FortiOS user databases and session tokens, crack hashes using Hashcat and Hashtopolis on rented GPUs, create rogue administrator accounts, and enumerate Active Directory. SOCRadar verified more than 86,644 compromised devices across 194 countries, and the access has served as an entry point for INC/Lynx and Payload ransomware affiliates. How can organizations test their exposure to the threats outlined in JCSA-20261006-01? SafeBreach customers can validate their defenses against FortiBleed by running the simulations mapped to JCSA-20261006-01. Customers can select CYBERSECURITY ADVISORY FortiBleed on the SafeBreach Scenarios page, filter the Attack Playbook by the Threat Name CYBERSECURITY ADVISORY FortiBleed, or open the Known Threat Series report and run the FortiBleed simulations. Results show whether existing controls detect and block the campaign’s brute force, password spraying, Active Directory discovery, exfiltration, and command-and-control communication. What should FortiGate administrators do first? FortiGate administrators should first remove or restrict internet-facing management, terminate all active administrative and VPN sessions, and reset all Fortinet VPN and administrator passwords. JCSA-20261006-01 also recommends requiring phishing-resistant MFA, confirming that administrator passwords use PBKDF2 instead of legacy SHA-256, removing unknown REST API keys, and comparing FortiGate user configurations against a known-good baseline. You Might Also Be Interested In Research Blog Gunra Ransomware: SafeBreach Coverage for CISA Alert AA26-222A Read More Solution Brief SafeBreach CTEM Platform Read More Solution Brief SafeBreach Propagate Read More Get the latest research and news

凭据窃取vpn事件报告 @safebreach-blog source ↗
Recommended8

解压路径未校验,恶意 3D 资产即可落地可执行文件,0day 未修补

(0Day) Genesis World decompress_usdz Directory Traversal Remote Code Execution Vulnerability October 13th, 2026 ZDI-26-752 ZDI-CAN-32131 CVE ID CVSS Score 7.8 AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H Affected Vendors Genesis World Affected Products Genesis World Vulnerability Details This vulnerability allows remote attackers to execute arbitrary code on affected installations of Genesis World. Interaction with this library is required to exploit this vulnerability but attack vectors may vary depending on the implementation. The specific flaw exists within the decompress_usdz function. The issue results from the lack of proper validation of a user-supplied path prior to using it in file operations. An attacker can leverage this vulnerability to execute code in the context of the current user. Additional Details 07/16/26 - ZDI emailed a Genesis World contributor to request a secure channel for vulnerability disclosures 07/28/26 - ZDI asked the vendor to update their security policy via a GitHub issue 07/28/26 - the vendor closed the issue 08/06/26 - ZDI followed up requesting the vendor’s PSIRT contacts 09/11/26 - ZDI notified the vendor of the intention to publish the case as a 0-day advisory -- Mitigation: Given the nature of the vulnerability, the only salient mitigation strategy is to restrict interaction with the product Disclosure Timeline 2026-09-25 - Vulnerability reported to vendor 2026-10-13 - Coordinated public release of advisory 2026-10-13 - Advisory Updated Credit Aaron Luo of VicOne LAB R7 Back to Advisories

rce路径穿越0daycve @zdi-advisories source ↗

AI 自动化侦察利用的宣称缺乏实证,检测工程需区分貌似合理与证据

Oct 12, 2026 Absence of Evidence: What an AI-Orchestrated Intrusion Harness Teaches Us About AI Detection Engineering SecFlow, an AI-powered orchestration framework that uses AI models to automate reconnaissance and exploitation, provided a real-world lesson about why plausible is not the same as proof. Read SafeBreach expert Adrian Culley’s thoughts on the take-away for defenders . Summary What happened: Hunt.io researchers documented SecFlow, an AI orchestration harness run by a Chinese-speaking operator (“nie”) that used Claude, Qwen, and DeepSeek as interchangeable workers for recon, exploitation, and reporting. The failure: One worker claimed a successful Apache Shiro exploit that was never verified. Because the harness treats prior output as ground truth, more than 27 follow-on tasks were built on it, and all of them failed. The capability is still real: The same harness dumped LSASS credentials from a Beijing district government environment, accessed 800+ database accounts, and generated steganographically concealed web shells. The infrastructure layer: Separate Lumen Black Lotus Labs research, not linked to SecFlow, describes an “Infrastructure Quartermaster” that rents reconnaissance (QScan) and relay networks (Fast Labyrinth) to multiple China-nexus groups. Attribution: Neither report names China’s Ministry of State Security. Targeting is consistent with state intelligence collection, but formal attribution isn’t established. The defender lesson: AI-generated detection rules have the mirror-image problem. A rule that loads without errors is not a rule that fires. Only live execution against your controls proves coverage. Author: Adrian Culley, Offensive Security Expert & Senior Sales Engineer “Absence of evidence is not evidence of absence.” Carl Sagan popularized the line, drawing on a much older statistical idea, and it’s one of the most misapplied sentences in science. It may also be the most important sentence for anyone putting AI anywhere near offensive or defensive security work. The use of AI in security is moving quickly from supporting the process to taking action. Defenders use it to speed up detection engineering: drafting detection rules , mapping attack techniques, and writing playbooks. Attackers use it too, and increasingly they aren’t just prompting a chatbot. They’re wiring several models together into automated pipelines, known as AI harnesses, that plan, execute, and report on an operation with minimal human review. That shift creates a new kind of risk for attackers and defenders alike. When a model says something worked, its output often looks exactly like success: well-formed, confident, and in the expected format. But looking right and being right are different things. A system that never checks the difference will keep building on mistakes it doesn’t know it made. Recent research into a China-nexus intrusion framework called SecFlow shows how this plays out in the real world. The harness got genuine access to a government network. It also spent effort on more than 27 follow-up tasks that were doomed from the start, because one earlier step claimed a success that nobody ever verified. The same failure is quietly showing up in defenders’ SOCs, where AI-generated detection rules can look ready to deploy without ever being proven to fire. To understand why, it helps to start with how an AI harness actually works. PREFER TO LISTEN? Check out episode 80 of The Cyber Resilience Brief, “The AI Harness That Never Checked Its Work,” where host Tova Dvorin is joined by SafeBreach expert Adrian Culley to unpack Hunt.io’s SecFlow research incident. Listen now . First, what is an AI harness? A large language model on its own answers one prompt at a time. An AI harness (sometimes called an orchestration framework or agent harness) is the software around the model that turns it into something that can complete multi-step work. A harness typically: Breaks a goal into tasks , often following a predefined workflow or playbook Routes each task to a model , sometimes spreading work across several models or several instances of the same one Gives models tools , like the ability to run commands, call APIs, scan networks, or read and write files Carries state between steps , so one task’s output becomes the next task’s input Decides what happens next based on what the previous step reported Harnesses are now common in legitimate work, from coding assistants to security automation. They’re what lets AI move from answering questions to carrying out a sequence of actions with little human involvement. That last point is also the weak spot. A harness is only as reliable as the results it passes along. If it accepts a model’s report of success without checking it, a single wrong answer can shape every step that follows. That’s exactly what happened recently with SecFlow. SecFlow: A Harness That Never Checked Its Own Homework Hunt.io’s research describes SecFlow as an AI orchestration framework built by a Chinese-speaking operator using the handle “nie.” SecFlow farmed out reconnaissance, exploitation, and reporting to Claude, Qwen, and DeepSeek instances, which ran as interchangeable workers. The architecture was straightforward. A master workflow file (a SKILL.md policy file) pushed each target through initialization, reconnaissance, exploitation, post-exploitation, and reporting. Workers shared a file system, so one worker’s output became the next worker’s input. That design is where things broke down. At one point, a worker reported that it had successfully exploited an Apache Shiro vulnerability. No one confirmed it. Because the harness treated the prior worker’s output as ground truth, the unverified success was carried into more than 27 subsequent tests, all of which failed because the original exploit had never worked. The system wasn’t lying, exactly. The output was plausible: a well-formed “exploitation successful” message in the shape of a good result. But no one, human or machine, went back to check what mattered. Did a shell actually open? Did a single byte of data leave the box? Plausible-looking success was treated as verified success. The Mirror Image: The Hidden Risks in AI Detection Engineering What is AI detection engineering? AI detection engineering is the use of large language models (LLMs) and other AI tools to build and maintain the rules that spot malicious activity in a security environment. Typical tasks include writing Sigma or SIEM queries, mapping detections to MITRE ATT&CK techniques, tuning thresholds, and drafting response playbooks. How does this incident relate to AI detection engineering? This isn’t only an offensive AI problem. It’s the same failure detection engineers face, seen from the defender’s side. Ask an LLM to write a Sigma rule for a specific attack technique and you’ll have one in about four seconds. It will almost certainly be valid YAML, its field mappings will look plausible, and it will probably load into your SIEM without a syntax error. None of that tells you whether it fires when the real technique runs against your real endpoint configuration. That’s because syntactically correct is not operationally correct. A rule that passes is not a rule that detects. AI-assisted detection engineering makes it easy to produce a large volume of rules that pass. That can create a false sense of coverage that is arguably worse than having no rule at all, because now someone believes they’re covered. “We have a rule for that” and “that rule fires when it should” are two very different sentences, and only the second describes a security control. This is where Sagan’s line cuts both ways. If a rule never fires because no one tests it against a live execution, that’s not evidence the technique isn’t happening in your environment. It’s evidence you never looked properly. The absence of an alert is not evidence of the absence of an attack. SecFlow’s Targets & Tradecraft Hunt.io researchers identified the SecFlow operator by the handle “nie,” which appeared consistently across a model service namespace, proxy credentials, and a family of domains hosting private relays for Claude, Qwen, and DeepSeek. The targeting spanned six countries, mostly government, political, and telecom organizations, including: A Fengtai District Government office automation environment in Beijing, China An attempted intrusion into the Kuomintang Party History Archive in Taiwan A web shell attempt against Indonesia’s Ministry of Foreign Affairs Port scanning against an Afghan telecom’s SMPP infrastructure That list reads like intelligence collection, not fraud. The tradecraft was real, too. SecFlow included a dedicated capability called Glutton for generating web shells in Java, .NET, and Node, delivered steganographically: XOR-encrypted payloads hidden in the RGB channels of PNG files to evade static signatures. The confirmed Fengtai compromise included LSASS credential dumping and extraction of 822 user account records from its OA database. The exploit chain covered eight named CVEs, including Shellshock, Log4Shell, Spring4Shell, Ghostcat, and the Apache Shiro deserialization flaw (CVE-2016-4437) behind the false success. The oldest, Shellshock, dates to 2014, yet these flaws are apparently still useful against unpatched estates in 2026. The Infrastructure Quartermaster: Recon & Relays as a Service Two separate pieces of research, published about a week apart in late summer 2026, describe what may be two halves of an emerging model. Neither report links the two operations. The first is Lumen Black Lotus Labs’ work on what they call the Infrastructure Quartermaster , a private technical entity whose administrative activity Lumen traced to Nanjing. It doesn’t carry out intrusions itself. It builds and rents out the plumbing, including: QScan: automated reconnaissance at scale Fast Labyrinth: an encrypted relay network running on legitimate commercial proxy subscriptions QT Proxy: an admin panel managing it all Multiple downstream Chinese threat groups appear to use this shared infrastructure. Instead of every crew building its own obfuscation network, a specialist quartermaster handles target discovery and traffic laundering for multiple customers. QScan profiled research universities (including aerospace and bioinformatics departments), government, defense, and critical infrastructure targets across the US, UK, Europe, and Asia-Pacific, rotating subnets on a 30-day cycle to avoid detection. Lumen tracked it for about a year before publishing. A Careful Word on Attribution Attribution isn’t for the faint-hearted, so let’s be precise. Neither report names China’s Ministry of State Security (MSS). Lumen describes the Quartermaster as China-nexus and its downstream users as Chinese threat groups. Hunt.io stops short of naming an intelligence service. There’s no government attribution statement and no MSS branding on the infrastructure. What we can say with more confidence is that the combined targeting (defense networks and research universities scanned by QScan; government systems, a party political archive, a foreign ministry, and telecom infrastructure targeted by SecFlow) looks like a state intelligence collection list, not a criminal one. The fairest framing: state-nexus, moderate confidence based on targeting pattern; formal intelligence-service attribution not established in public reporting. Dangerous & Sloppy at the Same Time SecFlow is real access and real waste. The harness that pulled LSASS dumps out of a government network also burned effort on more than 27 doomed tasks because of one unverified claim. The capability is real. The discipline inside it is not. That may be the most important strategic point in the story. People want AI-driven APT operations to be either terrifying and flawless or overhyped and harmless. The evidence says neither. It’s a real capability uplift for a state-nexus actor : faster reconnaissance, faster exploit customization, and faster web shell generation with evasion built in, all running inside a harness that inherits every epistemic weakness of the underlying models. That combination is what defenders need to plan for, not a caricature of either extreme. What Security Leaders Should Do Differently Whichever side you’re on, the takeaway is the same: don’t trust a system’s self-report of success. If you’re a detection engineer and an AI assistant hands you a rule, a mapping, or a playbook, treat it as a draft. It isn’t a verified control until you’ve executed the actual technique against it and watched it fire. If you’re red teaming with AI tooling , apply the same discipline in reverse: don’t let a model’s “exploit successful” claim flow into your next phase without confirming impact. Three Immediate Questions for SOC Leads Has it actually been executed? For every detection rule generated or substantially assisted by an LLM in the last six months, has anyone run the corresponding technique against it and confirmed that it fires? Not reviewed. Executed. Right now, this is one of the industry’s quiet problems. Are you separating “deployed” from “validated”? Is there a process gap where a rule sitting in the SIEM gets reported upward as “we’re covered”? When a detection stays silent, what do you conclude? When a rule stays silent in production, do you treat that as “we’re clean” or “we don’t know”? These are very different postures, and only one of them is honest. That third question is Sagan’s line in practice. A quiet SIEM isn’t proof of a quiet network. Validate Continuously, Not Once The defenders who get this right run continuous validation . They execute technique chains against their own controls on a recurring schedule instead of checking that a rule exists and looks reasonable. That’s the validation layer of Continuous Threat Exposure Management (CTEM) in practice, and it’s how SafeBreach customers close this gap : confirming that a rule fires against a live execution, not just that it compiles. The principle holds whatever platform does the testing. The question to take back to your team is: Can we prove, with evidence, that our top 10 detections fire against the real technique today, not six months ago when the rule was written? Ask the Same Question of Your AI Offensive Tooling If you’re evaluating AI tooling for offensive testing, red team automation, adversary emulation, or anything that generates and executes attack logic, ask it the same thing you’d ask a detection rule: Does it verify its own claimed outcomes, or does it just report what it was told to report? SecFlow shows what happens when the answer is the latter. Twenty-seven-plus wasted cycles is almost funny, until you remember that the same architecture pulled a real LSASS dump out of a real government network. What to Watch Over the Next 12 Months Whether harness architectures mature. If the next version of something like SecFlow adds a verification step, a worker that independently confirms exploitation before tasking follow-on work, the wasted-cycles problem disappears quickly. That would be a meaningful capability jump. Whether the Quartermaster model spreads. Shared, rented recon and relay infrastructure serving multiple downstream crews is a scaling pattern. If it works for this operator, others will copy it. Whether defensive AI gets the same scrutiny. Hunt.io gave us a rigorous failure-mode analysis on the offensive side. There’s comparatively little public scrutiny of whether AI-assisted detection engineering is quietly shipping the same unverified-success problem into SOCs. The Bottom Line Absence of evidence is not evidence of absence. That’s true for the state actor you haven’t formally attributed yet, and it’s true for the detection rule you’ve never actually fired. In both cases, the only honest answer to “are we covered?” is one you’ve already tested. SafeBreach runs real attack techniques against your security controls, safely and continuously, so you can see which detections fire, which ones fail silently, and where the gaps are. That includes detections written by your team and ones drafted with AI. Request a demo to see how SafeBreach helps security teams move from “we have a rule for that” to “we’ve tested it.” About the Author Adrian Culley is Head of Engineering, EMEA for SafeBreach. He has over 35 years of computer security experience, particularly specialising in security operations, artificial intelligence, offensive security, threat intelligence, and malware — in short, cyber resilience. He is a core member of the OWASP Large Language Model Top Ten risks working group. He is a graduate of the Information Security Group at Royal Holloway, University of London, and an occasional visiting lecturer there. Adrian was previously a Detective with Scotland Yard, working around the world with their Special Operations Directorate, and latterly served with the Computer Crime Unit there. He has also worked directly for the UK Central Government in a global role. He worked for Palo Alto Networks for five years as a Consulting Engineer and Head of Industry 4.0. Previously, he has also worked for PwC, Trellix, Cybereason, Guidance Software, AccessData, and Damballa. Adrian has a strong background in digital forensics and has worked with government, NGO, military, intelligence, law enforcement, and commercial clients around the world. Frequently Asked Questions What is AI detection engineering? AI detection engineering is the use of large language models (LLMs) and other AI tools to build and maintain the rules that spot malicious activity in a security environment. Typical tasks include writing Sigma or SIEM queries, mapping detections to MITRE ATT&CK techniques, tuning thresholds, and drafting response playbooks. AI can cut rule creation from hours to seconds. But a rule AI generates has only been checked for syntax, not proven to work. It becomes a real security control only after the matching attack technique has been run against it in your environment and the rule has fired. What is an AI harness? An AI harness is the software layer that wraps one or more large language models (LLMs) so they can complete multi-step work. It breaks a goal into tasks, assigns them to models, gives those models tools like command execution or network scanning, and passes each step’s output to the next. Harnesses power legitimate tools such as coding assistants and security automation, and threat actors now use them too. What is SecFlow? SecFlow is an AI orchestration harness documented by Hunt.io researchers. Built by a Chinese-speaking operator using the handle “nie,” it assigned reconnaissance, exploitation, post-exploitation, and reporting tasks to Claude, Qwen, and DeepSeek instances running as interchangeable workers, with each worker’s output feeding the next through a shared file system. Why did SecFlow waste more than 27 follow-on tasks? One worker claimed it had exploited an Apache Shiro vulnerability, but the claim was never verified. Because the harness treats prior output as ground truth, more than 27 subsequent tasks were built on that false success, and all of them failed. What is the “Infrastructure Quartermaster”? It’s Lumen Black Lotus Labs’ name for a private technical entity, likely based in Nanjing, that doesn’t carry out intrusions itself but rents infrastructure to multiple China-nexus threat groups . That includes QScan for automated reconnaissance, the Fast Labyrinth encrypted relay network, and the QT Proxy admin panel. Is SecFlow attributed to China’s Ministry of State Security? No. Neither the Hunt.io nor the Lumen reporting names the MSS. Lumen describes China-nexus infrastructure and Hunt.io a Chinese-speaking operator; the targeting in both is consistent with state intelligence collection. That supports a reasonable inference of state alignment, not a confirmed intelligence-service attribution. How does this relate to AI-assisted detection engineering? It’s the mirror-image failure. LLMs can produce syntactically valid detection rules in seconds, but a rule that loads into a SIEM isn’t proven to fire against the real technique in your environment. Treating plausible output as verified coverage creates a false sense of security. How can security teams validate AI-generated detection rules? Execute the corresponding attack technique against your controls and confirm the rule fires, then keep doing it on a recurring schedule. Continuous security validation , part of a CTEM program , separates rules that are deployed from rules that are proven. What should I ask vendors of AI-driven offensive testing tools? Ask whether the tool independently verifies its claimed outcomes, such as confirming that a shell opened or data actually left the system, or whether it simply reports what it was told to report. You Might Also Be Interested In Podcast Videos Podcast: Anthropic’s CISO Guide to Agentic AI: Your Next Insider Threat Is an AI Agent Listen Now Partner Solution Brief SafeBreach & Anvilogic Read More China State-Sponsored Cyber Threat Actors: A Comprehensive Guide Read More Get the latest research and news

ai_agentllm报告检测 @safebreach-blog source ↗

服务端解析模型即被用作请求跳板,可用于内网探测,暂无补丁

(0Day) Genesis World glTF Server-Side Request Forgery Vulnerability October 13th, 2026 ZDI-26-755 ZDI-CAN-32132 CVE ID CVSS Score 6.1 AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:L/A:N Affected Vendors Genesis World Affected Products Genesis World Vulnerability Details This vulnerability allows remote attackers to initiate arbitrary server-side requests on affected installations of Genesis World. Interaction with this library is required to exploit this vulnerability but attack vectors may vary depending on the implementation. The specific flaw exists within the parsing of glTF files. The issue results from the lack of proper validation of a URI prior to accessing resources. An attacker can leverage this vulnerability to initiate requests in the context of the current process. Additional Details 07/16/26 - ZDI emailed a Genesis World contributor to request a secure channel for vulnerability disclosures 07/28/26 - ZDI asked the vendor to update their security policy via a GitHub issue 07/28/26 - the vendor closed the issue 08/06/26 - ZDI followed up requesting the vendor’s PSIRT contacts 09/11/26 - ZDI notified the vendor of the intention to publish the case as a 0-day advisory -- Mitigation: Given the nature of the vulnerability, the only salient mitigation strategy is to restrict interaction with the product Disclosure Timeline 2026-09-25 - Vulnerability reported to vendor 2026-10-13 - Coordinated public release of advisory 2026-10-13 - Advisory Updated Credit Aaron Luo of VicOne LAB R7 Back to Advisories

ssrf0daycve @zdi-advisories source ↗

已有低权账号即可越权执行命令,7.5 全系列受影响且无虚拟补丁。

PSIRT Broken authorization Summary A missing authorization vulnerability [CWE-862] in FortiSIEM GUI may allow an authenticated attacker to execute unauthorized code or commands via Specially crafted HTTP requests Version Affected Solution FortiSIEM 7.6 Not affected Not Applicable FortiSIEM 7.5 7.5 all versions Migrate to a fixed release Acknowledgement Discovered and Responsibly reported by external researcher khalooda0x - Khaled ibn Al-Walid. Virtual Patch No Timeline 2026-10-13: Initial publication IR Number FG-IR-26-183 Published Date Oct 13, 2026 Component GUI Severity Low Discovered External Attack Type Authenticated Known Exploited No CVSSv3 Score 3.4 Impact Execute unauthorized code or commands CVE ID CVE-2026-106176 Download CVRF CSAF

授权绕过固件cve @forti-ir source ↗

未认证即可打瘫管理界面,8.0.0 与 7.6.4–7.6.6 受影响,可先限制管理口暴露。

PSIRT GUI Denial of Service Summary An Improper Locking vulnerability [CWE-667] in FortiOS may allow an unauthenticated attacker with network access to the device to cause a denial of service of the HTTPS management interface via crafted HTTP requests. Version Affected Solution FortiOS 8.0 8.0.0 Upgrade to 8.0.1 or above FortiOS 7.6 7.6.4 through 7.6.6 Upgrade to 7.6.7 or above FortiOS 7.4 Not affected Not Applicable FortiOS 7.2 Not affected Not Applicable Follow the recommended upgrade path using our tool at: https://docs.fortinet.com/upgrade-tool Virtual Patch named "FG-VD-61291.0day." is available in FMWP db update 26.073 Acknowledgement Fortinet is pleased to thank Oded Weissman, and clxhzg with Tu Longfeng for reporting this vulnerability under responsible disclosure. Virtual Patch Yes Timeline 2026-10-13: Initial publication IR Number FG-IR-26-181 Published Date Oct 13, 2026 Component GUI Severity Medium Discovered External Attack Type Unauthenticated Known Exploited No CVSSv3 Score 6.7 Impact Denial of service CVE ID CVE-2026-106174 Download CVRF CSAF

dos固件cve @forti-ir source ↗

已认证即可注入并执行命令,7.2/7.4/7.6 均提供修复版本,邮件网关属于高价值目标。

PSIRT SQL command injection in webmail Summary An improper neutralization of special elements used in an SQL Command ("SQL Injection") vulnerability [CWE-89] in FortiMail webmail may allow an authenticated attacker to execute unauthorized code or commands via specifically crafted HTTP requests. Version Affected Solution FortiMail 8.0 Not affected Not Applicable FortiMail 7.6 7.6.0 through 7.6.3 Upgrade to 7.6.4 or above FortiMail 7.4 7.4.0 through 7.4.5 Upgrade to 7.4.6 or above FortiMail 7.2 7.2.0 through 7.2.8 Upgrade to 7.2.9 or above Fortinet in Q1/26 has remediated this issue in FortiMail Cloud and hence the customers need not perform any action. Acknowledgement Internally reported and discovered by Jaguar Perlas of Burnaby Infosec team. Virtual Patch No Timeline 2026-10-13: Initial publication IR Number FG-IR-26-184 Published Date Oct 13, 2026 Component GUI Severity High Discovered Internal Attack Type Authenticated Known Exploited No CVSSv3 Score 7.7 Impact Escalation of privilege CVE ID CVE-2025-57739 Download CVRF CSAF

sqlircecve @forti-ir source ↗

需先持有集群节点证书与私钥才能冒充,门槛高但命中即命令执行,8.0.0 受影响。

PSIRT arbitrary command execution in fgfmsd daemon Summary An improper neutralization of special elements used in an OS command ('OS Command Injection') vulnerability [CWE-78] in FortiManager fgfmd daemon may allow a remote attacker impersonating a fortimanager cluster node to execute unauthorized code or commands via specially crafted packets. Impersonation of a Fortimanager cluster node requires possessing the node's certificate and private key. Version Affected Solution FortiManager 8.0 8.0.0 Upgrade to 8.0.1 or above FortiManager 7.6 7.6.0 through 7.6.6 Upgrade to 7.6.7 or above FortiManager 7.4 Not affected Not Applicable FortiManager 7.2 Not affected Not Applicable FortiManager Cloud 8.0 Not affected Not Applicable FortiManager Cloud 7.6 7.6.2 through 7.6.6 Upgrade to 7.6.7 or above FortiManager Cloud 7.4 Not affected Not Applicable FortiManager Cloud 7.2 Not affected Not Applicable Virtual Patch No Timeline 2026-10-13: Initial publication IR Number FG-IR-26-177 Published Date Oct 13, 2026 Component OTHERS Severity Medium Discovered External Attack Type Authenticated Known Exploited No CVSSv3 Score 5.3 Impact Escalation of privilege CVE ID CVE-2026-106171 Download CVRF CSAF

命令注入rcecve @forti-ir source ↗

订阅授权接入本地智能体的官方协议与限制,并含可直接运行的参考实现。

开发工具 AI 产品与平台 很多人每个月都在给ChatGPT Plus或Pro付费,但写代码或使用本地智能体时,用的开发工具往往还要再填一个独立的API key,按token消耗在另一张账单上扣款。既然订阅里已经包含了充足的推理算力,为什么不能直接把这个额度接到自己的开发环境里? 过去要想用订阅额度调用第三方工具,只能借用官方客户端公开的身份标识,在官方没有明确禁止的灰色地带调用私有接口。直到2026年9月,OpenAI推出了Sign in with ChatGPT,这件事才有了官方的协议。这篇文章主要想介绍把ChatGPT订阅接入外部工具的技术演进、官方协议的具体规则和限制,以及在官方API key、官方订阅授权和复用官方客户端订阅身份这三种做法之间怎么取舍。文章的末尾我们也会提供一个自己实现的、可以直接运行的参考实现。 两条账单:API 为什么要按 token 另付一遍钱 如果你平时在本地折腾 AI 工具链,对这种两张账单并行的尴尬应该不陌生;团队和公司把用量放大之后,这笔重复开销只会更扎眼:一边是信用卡每月固定扣款的 ChatGPT 订阅费,一边是 OpenAI 开发者后台每隔几天就自动充值的 API 账单。调用的明明是同一家公司的模型,为什么手头现成的订阅额度,就不能用来抵扣写代码和跑智能体的消耗? 这两笔账为什么不通,得从它们各自的收钱方式说起。面向开发者的开放平台,核心是 Responses API 推理端点 ,它纯粹按实际输入和输出的 token 数量计费。翻看 官方最新定价标准 ,专用模型 gpt-6.1-sol ,每 100 万 token 输入 2 美元、输出 10 美元;换成更高一档的旗舰模型 gpt-6-astra ,输出要涨到 50 美元。中间还有轻量模型和折中档,单价便宜很多,但本地工具只要频繁把大代码库塞进上下文、让模型跑多轮长思考,token 累积起来开销涨得很快。 另一面是面向普通用户或团队的 ChatGPT 订阅 。个人从免费的 Free、轻量的 Go,到主力档 Plus(每月 20 美元)都有,再到 按算力分档的 Pro 。团队则按人头计费, ChatGPT 商业版方案 标准坐席按月付每人每月 25 美元,年付有折扣。 还有一点容易忽略:ChatGPT 订阅本来就内置了 Codex(官方的编程功能),而且 所有套餐的 Codex 和 ChatGPT Work 共用一个用量池 。这块额度折算成 API 调用相当可观, SemiAnalysis 的测算 认为,OpenAI 订阅的 API-equivalent value 大约是订阅费的 12 倍,以 $200 档的 Pro 换算下来接近两三千美元的 API 花费。换句话说,你手里本来就有一大笔没用起来的算力,却还在为第三方工具另开一张 API 账单,这才让想用订阅这件事变得很有吸引力。 这两套产品在计费和协议体系上一直是分开的。官方在商业版的 FAQ 里也明确说明,ChatGPT 商业版订阅不包括 API 用量,API 使用需要额外付费。从协议上看,消费级产品由 OpenAI 使用条款 管辖,但开发者 API 用的是 商业服务协议 ,商业条款中明确指出不适用于消费级个人服务。两边账户余额也不互通。很多人有高级订阅却在第三方工具里还要绑卡,就是因为官方在商业规则上把这两个体系隔开了。 想用订阅跑自己的工具,过去只有一条灰路 面对这两张账单,一个自然的想法就是能不能让外部工具假扮成官方客户端,直接去扣订阅里的额度。这里面主要是两件事:先是身份认证,确认当前是谁在发起调用;再是计费归属,确定这笔调用由谁来买单。 API key 可以轻松实现计费,但用户需要拥有开发者账号并且持续充值。ChatGPT 的订阅虽然预付了算力,但官方只支持他们的网页端、移动端和桌面客户端。不过官方自己的编程助手 Codex 是个例外。Codex 客户端(前面提到的 gpt-6.1-sol 就是它默认用的代码模型)提供了两种平行的登录途径:传统 API key,和 基于 ChatGPT 订阅的登录认证 。 既然官方的 Codex 客户端已经支持用 ChatGPT 订阅来登录,那其他第三方客户端有没有可能也用同样的方式呢?遗憾的是,官方并没有提供 API。因此,开源社区通过逆向工程官方客户端,发现了整个实现流程。正如 开发者社区的深入讨论 所揭示的那样,不少编辑器里的代码插件(比如 Roo Code)选择直接复用官方 Codex 客户端公开发布的 OAuth 客户端标识符 client_id,在本地调起浏览器完成授权,接着把推理请求直接转发到 ChatGPT 的私有后端端点 chatgpt.com/backend-api/codex/responses 。 这种做法技术上确实有效,用户不用多花钱,就能用上订阅里的 Codex 算力。但在契约性质上,它从来没有得到过官方的正式支持。 官方的态度一直是暧昧的。2026 年 2 月 9 日,OpenAI 负责 Codex 的维护者 etraut-openai 在 GitHub 讨论区公开回应 开发者时表达了个人看法:作为工程师而非律师,官方的使用条款和代码开源协议相对宽松,像开源的终端编程工具 OpenCode 也在做类似的事。这段非正式的回复,不少开发者当时把它当成了一种事实上的默许。 但在官方的文档和协议中,这个领域始终是模棱两可的灰色地带。官方的使用条款明确禁止与他人共享账户凭据,也明文禁止绕过任何速率限制、技术保护或安全缓解措施。 OpenHands 软件代理开发团队 在进行了合规审计后也指出,主流模型厂商明确表示不鼓励甚至禁止使用第三方客户端对消费级产品进行自动化调用,对于生产环境中的严肃需求,他们仍然建议通过官方 API key 的方式进行调用。 这种局面一直维持到 2026 年秋季:大家虽然靠着私有接口把工具跑通了,但心里清楚这是一条未承诺稳定、缺乏 SLA 保障且随时可能因接口变动而失效的临时通路。 官方路径:Sign in with ChatGPT 2026 年 9 月 29 日,OpenAI 在 DevDay 上正式发布了 Sign in with ChatGPT 服务 。它要解决的核心问题,是让第三方工具在用户授权之后,可以名正言顺地花掉用户自己 ChatGPT 订阅里的额度。这是官方第一次公开承认并放行这种用法,也是开源项目和本地工具摆脱私有端点的正规入口。首发就公布了 16 家合作伙伴。 Sign in with ChatGPT 把授权拆成两层:确认身份的身份层,和真正花掉订阅额度的额度层,中间用一个可选开关隔开 在官方公布的生态体系中,开源软件获得了明确的席位。根据 官方接入指南与集成名单 ,首批开源集成伙伴涵盖了 OpenClaw、OpenCode、Pi 以及 T3 等项目;商业级试用应用则包含了 Devin、Notion、Vercel、Warp、Airtable、GitLab 与 Supabase 等产品。 这引入了一个关键的新概念:授权分为两个层次: 1. 基础身份认证层 :在 官方快速入门文档 中定义,其授权范围为标准的 openid profile email 。该层仅负责确认用户的 ChatGPT 账号身份并返回 ID token,官方在此处做出明确说明:基础身份权限不提供对用户 ChatGPT 历史对话或 OpenAI API 资源的任何访问能力。 2. 订阅额度共享层 :在 面向开源项目的额度共享规范 中定义,要求显式申请 offline_access resource.invoke chatgpt.tokens.use.direct 权限,同时必须把 resource 参数设为 https://api.openai.com/v1 。只有用户在授权界面勾选允许该应用调用自己的订阅额度,外部工具才能发起实际的模型推理。 在开放范围上,商业应用的接入仍处于有限试用阶段;但对于订阅额度调用能力,官方规定在发布初期即全面向所有开源项目、在本地运行的个人项目以及选定的私有客户端开放。 在所有权和计费方面, 官方文档 指出,拥有Go、Plus或Pro订阅的用户可以自行授权兼容的应用程序使用自己的订阅,无需生成或共享个人API key。所有合格请求都会直接计入用户套餐里包含的用量额度。这是一种典型的用户自带订阅模式(Bring Your Own, BYO),不用大家混用同一个共享账号。用户始终可以在 ChatGPT 的设置中心随时查看已授权工具的活跃状态,监控算力使用情况,并随时单方面撤销授权。 对于开源或本地部署的项目,官方提供了一套动态注册机制,不需要提前申请客户端密钥。客户端首次向 账户授权端点 发起请求时,统一将 client_id 设置为公共预留标识 dynamic_agent_client ,并传入当前运行宿主的唯一标识 ext_agent_host_id 以及工具名称提示 agent_name_hint 。 授权流程在本地环回端点 127.0.0.1:1455/auth/callback 接收回调。验证通过后,OpenAI 会在响应中下发一个绑定当前用户账号与所选 workspace 的正式 client_id ,供后续流程长期复用。整个过程既不需要存储客户端密钥,也不需要开发者注册商业 API key。 在推理端点上,官方给出了很明确的技术约束:所有推理请求必须通过 POST 发送到公开端点 api.openai.com/v1/responses 。官方文档以加粗字体发出警告:必须使用上述公开 Responses API 端点,严禁将请求导向 ChatGPT 的 backend-api 私有端点。 每次推理请求必须在 JSON 载荷中强制声明 store:false (不保存用于后续检索或续接的响应状态)与 stream:true (强制流式输出)。凭证有效周期方面,访问令牌的有效时间为 1 小时,刷新令牌则支持 30 天滑动窗口轮转。 官方将当前订阅额度共享定位为预览阶段,相比于完全自由的 API key,它仍然包含很多严格的限制。根据 预览阶段限制说明 ,调用方无法使用常见的采样调优参数,诸如 temperature 、 max_output_tokens 、 prompt 、 background 、 conversation 、 metadata 以及 moderation 等字段均不支持配置;不能使用官方提供的很多功能,比如内置的代码解释器、文件检索、图像生成、原生计算机操作、云端托管的 MCP 连接器等;不能传入音视频作为输入、不能进行文件上传和语音转写等。此外,这个接口也不支持基于服务端保存的历史自动进行对话延续,而是在工具端手动把历史消息加入请求中。 一个能跑的参考实现:chat_gpt_oauth skill 上一节介绍的是官方给的正道。这一节要看的是一条更早、也更直接、但已经不推荐长期依赖的路:把前面提到的那条灰路写成一个能跑的最小实现。它更像一份教学样本,价值在于把用订阅换 token 的整套机制摊开给你看清,本身不适合当作长期依赖。我们把它开源在了 chat-gpt-oauth-skill 代码库 。 要说明的是,它带一个能单独跑的命令行,但你完全可以不用它,那个命令行本身意义不大。它真正的定位是一份 reference implementation:把它集成进你自己的 agent 或 sub-agent 里,让用 ChatGPT 订阅发请求这件事,变成工具链里的一个模块。它完整演示了 OAuth 凭据交换、令牌生命周期管理和 Codex 推理请求构造,读代码就能看清 subscription token 长什么样、请求是怎么发出去的。 安装不需要自己啃文档。把仓库链接直接丢给你的 AI 编程工具(Codex、Claude Code、OpenCode 都行),让它先把仓库读一遍,再照着里面的说明安装、配置、跑通 demo 就行。它默认把 token 以明文 JSON 存在 ~/.chatgpt_oauth/token.json (权限 0600),这是刻意留给调试观察的教学设计,不是生产级的安全做法。 它走的通信目标是官方 Codex 私有后端,和上一节的官方 Sign in with ChatGPT 是同一诉求下的两条路径:官方协议打 api.openai.com ,这个 skill 则复用官方 Codex 客户端的既有身份打私有端点。所以它不是一个官方 SDK,依赖的是没有 SLA 保证的私有端点,只适合个人技术探索、学习验证,以及完全由你自己掌控的本地轻量工具。 总结:还有哪些路,以及怎么选 把这些路并排放在一起,才好看清每一条适合谁。 接入 ChatGPT 的三条路:API key 按量付费、Sign in with ChatGPT 用订阅额度、复用 Codex 身份走灰色地带,各自代价不同 目前接入 OpenAI 模型的路径主要有三条: 第一条是标准的 官方 API key 路线 。它是 OpenAI 开发者生态中的一等公民,打向标准的 api.openai.com/v1/responses 端点,受官方 Services Agreement 管辖。这条路径的能力最完整:支持所选模型开放的全部参数、官方的代码解释器与文件检索等全套工具链、各类模态输入。它的代价是按 token 计费,无法共享个人 ChatGPT 订阅内已有的额度。 第二条是官方的 Sign in with ChatGPT 订阅额度路线。它专为开源项目和自用脚本提供官方认可的订阅消费通道,用户可以光明正大地携带自己的 Plus、Pro 订阅登录,而无需在第三方工具中暴露高风险的 API key,并且可以随时在 ChatGPT 设置中查看与切断授权。但它的代价是目前仍处于预览阶段,官方裁掉了不少进阶微调参数和托管工具,输入也只支持所选模型允许的文本、图像和文件。 第三条是直接复用 Codex 客户端身份的逆向协议路线(如 chat_gpt_oauth skill),它在本地可以最大限度地压榨出订阅所包含的 Codex 专用模型能力,但它运行在官方未明文背书的灰色地带,缺乏长期稳定性保证,仅适合个人折腾与本地过渡。 这三条里,一条是另开 API 账单,两条是想办法用上订阅额度。 面对这些分支,开发者的选型判据可以归纳为以下几条: - 追求零额外成本,希望消耗个人闲置订阅 :优先尝试 Sign in with ChatGPT;如果受限于预览期的参数限制或需要特定的 Codex 交互体验,且完全清楚其非官方性质,可暂时采用类似 chat_gpt_oauth 的本地参考实现。 - 追求生产级稳定性,需要复杂的工程调优与多模态工具 :直接选用官方 API key。为按量付费买单,换取的是更完整的官方 API 能力与商业协议保障。 使用个人订阅去驱动第三方工具而非按量计价的 API,在过去很长一段时间里只能寄希望于借用官方客户端身份在灰色地带探索。而 2026 年 9 月推出的 Sign in with ChatGPT,首次在官方契约层面为开源与本地开发打开了正规入口。尽管目前这条正规通道仍然受到严格的预览限制,但它标志着消费级订阅与第三方开发者工具之间的割裂开始弥合。 最后补一句和主线无关的:如果你要解决的是另一类问题,比如让程序在云端长时间自主运行,官方另有 Agents API 。它把驱动 Codex 的长任务编排能力做成托管服务开放出来,但按标准 API 费率计费、不消耗订阅额度,这里就不展开了。

llmoauthai_agent工具 @yage-share source ↗
More12

(0Day) Genesis World URDF Directory Traversal Information Disclosure Vulnerability October 13th, 2026 ZDI-26-754 ZDI-CAN-32130 CVE ID CVSS Score 5.5 AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:N/A:N Affected Vendors Genesis World Affected Products Genesis World Vulnerability Details This vulnerability allows remote attackers to disclose sensitive information on affected installations of Genesis World. Interaction with this library is required to exploit this vulnerability but attack vectors may vary depending on the implementation. The specific flaw exists within the parsing of URDF files. The issue results from the lack of proper validation of a user-supplied path prior to using it in file operations. An attacker can leverage this vulnerability to disclose information in the context of the current process. Additional Details 07/16/26 - ZDI emailed a Genesis World contributor to request a secure channel for vulnerability disclosures 07/28/26 - ZDI asked the vendor to update their security policy via a GitHub issue 07/28/26 - the vendor closed the issue 08/06/26 - ZDI followed up requesting the vendor’s PSIRT contacts 09/11/26 - ZDI notified the vendor of the intention to publish the case as a 0-day advisory -- Mitigation: Given the nature of the vulnerability, the only salient mitigation strategy is to restrict interaction with the product Disclosure Timeline 2026-09-25 - Vulnerability reported to vendor 2026-10-13 - Coordinated public release of advisory 2026-10-13 - Advisory Updated Credit Aaron Luo of VicOne LAB R7 Back to Advisories

路径穿越cve0day @zdi-advisories source ↗

(0Day) Genesis World MJCF Directory Traversal Information Disclosure Vulnerability October 13th, 2026 ZDI-26-753 ZDI-CAN-32129 CVE ID CVSS Score 5.5 AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:N/A:N Affected Vendors Genesis World Affected Products Genesis World Vulnerability Details This vulnerability allows remote attackers to disclose sensitive information on affected installations of Genesis World. Interaction with this library is required to exploit this vulnerability but attack vectors may vary depending on the implementation. The specific flaw exists within the parsing of MJCF files. The issue results from the lack of proper validation of a user-supplied path prior to using it in file operations. An attacker can leverage this vulnerability to disclose information in the context of the current process. Additional Details 07/16/26 - ZDI emailed a Genesis World contributor to request a secure channel for vulnerability disclosures 07/28/26 - ZDI asked the vendor to update their security policy via a GitHub issue 07/28/26 - the vendor closed the issue 08/06/26 - ZDI followed up requesting the vendor’s PSIRT contacts 09/11/26 - ZDI notified the vendor of the intention to publish the case as a 0-day advisory -- Mitigation: Given the nature of the vulnerability, the only salient mitigation strategy is to restrict interaction with the product Disclosure Timeline 2026-09-25 - Vulnerability reported to vendor 2026-10-13 - Coordinated public release of advisory 2026-10-13 - Advisory Updated Credit Aaron Luo of VicOne LAB R7 Back to Advisories

路径穿越cve0day @zdi-advisories source ↗

Please enable cookies. Sorry, you have been blocked You are unable to access unknowncheats.me Why have I been blocked? This website is using a security service to protect itself from online attacks. The action you just performed triggered the security solution. There are several actions that could trigger this block including submitting a certain word or phrase, a SQL command or malformed data. What can I do to resolve this? You can email the site owner to let them know you were blocked. Please include what you were doing when this page came up and the Cloudflare Ray ID found at the bottom of this page. Cloudflare Ray ID: a4a261d58b00cb5f • Your IP: Click to reveal 52.238.24.41 • Performance & security by Cloudflare

内核逆向驱动 @unknowncheats-ac source ↗

Please enable cookies. Sorry, you have been blocked You are unable to access unknowncheats.me Why have I been blocked? This website is using a security service to protect itself from online attacks. The action you just performed triggered the security solution. There are several actions that could trigger this block including submitting a certain word or phrase, a SQL command or malformed data. What can I do to resolve this? You can email the site owner to let them know you were blocked. Please include what you were doing when this page came up and the Cloudflare Ray ID found at the bottom of this page. Cloudflare Ray ID: a4a261ed9f62cb5f • Your IP: Click to reveal 52.238.24.41 • Performance & security by Cloudflare

逆向议题 @unknowncheats-ac source ↗

Oct 9, 2026 Integrity Technology Group: SafeBreach Coverage for CISA Alert AA26-281A Learn how SafeBreach maps AA26-281A to simulations that test your exposure to the campaign’s activities being used to target and steal sensitive data from organizations worldwide, including U.S. critical infrastructure sectors. Summary Joint Cybersecurity Advisory AA26-281A, published October 8, 2026 and led by the FBI with CISA, the NSA, and seven international partners, details how Chinese government-linked threat actors enabled by the Integrity Technology Group combine automated scanning, large-scale botnets, and hands-on exploitation to steal sensitive data from organizations worldwide. The threat actors use TTPs consistent with activity publicly tracked as Flax Typhoon, Ethereal Panda, and Red Juliett. They scan for exposed services with open source tooling and a custom platform containing over 1,300 penetration testing scripts, gain access through cross-site scripting and password spraying against Microsoft Exchange, establish persistence by installing legitimate SoftEther VPN clients renamed to mimic Windows binaries, and exfiltrate mailboxes and Active Directory credentials with purpose-built scripts. The advisory confirms eight successfully exploited CVEs, five of which are newly added to CISA’s Known Exploited Vulnerabilities Catalog. This post outlines the campaign’s TTPs mapped to MITRE ATT&CK, summarizes the advisory’s IOCs, and covers the available SafeBreach simulations that validate whether your controls detect and block the campaign’s attacks, which involve reconnaissance and web application exploitation followed by credential access, masquerading, and covert exfiltration. Author: Noam Sagiv, Product Manager On October 8, 2026, the Federal Bureau of Investigation (FBI), Cybersecurity and Infrastructure Security Agency (CISA), National Security Agency (NSA), United Kingdom National Cyber Security Centre (NCSC-UK), Australian Signals Directorate’s Australian Cyber Security Centre (ASD’s ACSC), the Canadian Centre for Cyber Security (Cyber Centre), Japan’s National Police Agency (NPA) and National Cybersecurity Office (NCO), New Zealand’s National Cyber Security Centre (NCSC-NZ), and Spain’s Centro Nacional de Inteligencia (CNI) released a joint Cybersecurity Advisory (CSA) titled “Chinese Government-linked Cyber Threat Actors Combine Automated and Hands-on Hacking Tools to Steal Sensitive Data” (AA26-281A) . This advisory provides critical intelligence on the tactics, techniques, and procedures (TTPs) and indicators of compromise (IOCs) recovered from multiple FBI investigations into the Integrity Technology Group and the threat actors it enables. The threat actors have targeted victims across US critical infrastructure sectors, including Government Services and Facilities, Critical Manufacturing, Healthcare and Public Health, and Information Technology, as well as US law enforcement, education, and religious organizations and victims across Southeast Asia, Africa, and North America. What is the Integrity Technology Group Threat? Integrity Technology Group (Integrity Tech) is a China-based for-profit company with links to the Chinese government that the US Treasury Department sanctioned in January 2025 for its role in intrusions against US victims. According to the authoring organizations, Integrity Tech employs individuals who support malicious cyber activity in several ways: acquiring or building offensive cyber tools for use and sale, acquiring and hosting infrastructure, and directly compromising victim networks. The services Integrity Tech provides feed the broader Chinese cyber ecosystem, which aims to exfiltrate sensitive data from victims worldwide. The threat actors enabled by Integrity Tech use TTPs consistent with the cyber activity publicly known as Flax Typhoon, Ethereal Panda, and Red Juliett, among others. The advisory notes an important caveat: cybersecurity companies use different methods of tracking and attributing cyber actors, and these may not map 1:1 to the US government’s own methodology. The authoring organizations also note that these threat actors may perform activity unassociated with Integrity Tech. The activity is long-running. The threat actors have used the MicroScan penetration testing platform as early as 2017, and have been gaining access to victim networks and cloud services since at least mid-January 2021. What distinguishes this campaign is not novel malware but the industrial pairing of two modes of operation: large-scale automated scanning and botnet infrastructure that identifies soft targets at volume, and hands-on operators who then exploit what the automation finds. That botnet infrastructure has a track record: in September 2024, the FBI, NSA, and international partners reported that Integrity Tech controlled a botnet of more than 260,000 compromised routers and IoT devices, which researchers at Lumen’s Black Lotus Labs tracked as Raptor Train and assessed was likely operated by Flax Typhoon. The actors lean heavily on open source tooling and legitimate software—living-off-the-land (LOTL) techniques, commodity scanners pulled from GitHub, and an open-source VPN client—which keeps their footprint close to the noise floor of normal enterprise activity. The objective throughout is data theft, specifically email. The FBI recovered an archived email database used to target victim organizations, and observed the actors restricting access to exfiltrated data to IP addresses originating in Xiamen, China. Recent PRC-linked advisories SafeBreach has covered include Countering Chinese State-Sponsored Espionage Campaigns (AA25-239A) and Volt Typhoon (AA24-038A) . For the full set of government guidance, see CISA’s People’s Republic of China Cyber Threat page. Integrity Technology Group TTPs Mapped to MITRE ATT&CK The advisory maps Integrity Technology Group’s activity to MITRE ATT&CK for Enterprise, version 19 . Reconnaissance Integrity Technology Group uses a range of open source scanning tools to find vulnerabilities in networks and web applications [ T1595.002 ], including BBScan, dirsearch, Fscan, ksubdomain, masscan, Nmap, OneForAll, ShuiZe, and WPScan. They concentrate on ports 21 (FTP), 22 (SSH), 53 (DNS), 80 (HTTP), 443 (HTTPS), and 1080 (SOCKS), and when scanning websites with dirsearch they attempt to enumerate PHP and ASP (.NET) pages. The reliance on commodity GitHub tooling suggests the actors are hunting for the most vulnerable targets rather than pursuing specific victims. Since at least 2017, the actors have also operated MicroScan, a Python-based web application containing over 1,300 penetration testing scripts written to scan websites for specific vulnerabilities. The scripts target services including OpenSSL, Oracle WebLogic Server, Rejetto HFS, WordPress, Juniper ScreenOS, Jenkins, and Apache Struts. Initial Access Integrity Technology Group executes cross-site scripting (XSS) attacks that leverage unwitting third-party applications to modify content on a webpage [ T1189 ]. The FBI recovered one such payload: when executed on a vulnerable website, it modifies the page to display username and password fields, harvesting credentials from users who enter them. After submission, the payload generates a link to a password-protected .zip archive whose contents are decoded from bytes embedded in the payload itself and which contains the executable live700_v1.exe. Execution Access to victim networks and cloud services comes primarily through command line utilities built on exploit code written in Python [ T1059.006 ] and Go. Integrity Technology Group uses JavaScript [ T1059.007 ] and HTML to execute the XSS attacks described above. SoftEther installers are downloaded onto Windows victim devices using PowerShell and LOTL binaries [ T1059.001 ], and the Linux-based office-cli utility [ T1059.004 ] continuously targets and accesses Microsoft 365 accounts to automate mail exfiltration using configuration files including client_id, tenant_id, and secret. Integrity Technology Group also uses EBurst, an open source Python tool, to target Microsoft Office 365 accounts. EBurst compromises email accounts on Microsoft Exchange servers through multiple interfaces: Exchange Control Panel (ECP), Exchange Web Services (EWS), Offline Address Book (OAB), Outlook Web Access (OWA), Remote Procedure Call (RPC), API, MAPI, PowerShell, Autodiscover, and Microsoft-Server-ActiveSync. The authoring organizations specifically recommend that network defenders cover all of these interfaces when defending against EBurst. Persistence To establish persistence, Integrity Technology Group installs virtual private network (VPN) software clients on victim devices [ T1133 ] to obfuscate command and control (C2) communications and frustrate attribution. They typically download SoftEther installers from actor-controlled infrastructure using PowerShell or LOTL binaries on Windows, and curl or wget on Linux/Unix, then configure the client to reconnect automatically on startup. The advisory notes that endpoint detection software is less likely to flag SoftEther because it is legitimate VPN software. SoftEther has been a Flax Typhoon signature since Microsoft first named the group in 2023 . Defense Evasion Integrity Technology Group runs executables under the names of known, legitimate Windows software to evade detection [ T1036.003 ]. Observed examples include naming SoftEther installers conhost.exe or dllhost.exe, and the live700_v1.exe payload spawning a process named DiagTrack.exe, which then establishes encrypted communications over HTTP with the domain dns.studiocloud[.]xyz—a domain the FBI attributes to Integrity Tech. Credential Access Integrity Technology Group uses DC.exe to execute the DCSync replication technique [ T1003.006 ], copying account credentials, group membership details, and trust relationships out of Active Directory. DC.exe binds to the victim’s domain controller over RPC and uses the Directory Replication Service to retrieve a handle to the local security policy object, then queries the DNS domain name, domain security identifier (SID), and domain replication epoch. Against Microsoft Exchange, the actors use eburst.py to conduct both password spraying [ T1110.003 ] and password guessing [ T1110.001 ] to compromise email accounts. Collection Integrity Technology Group created a bot using the PHP script Curlc4.txt to obtain victim emails. The script interfaces with the Microsoft EWS API, giving the actors access to email and content items such as calendars and contacts [ T1114.002 ]. Before exfiltration the bot compresses emails and, in some instances, encrypts them using RC4 or AES-128-CBC [ T1560.003 ]. When no root directory is supplied as an argument, the script searches for a writeable directory to stage data in [ T1074.001 ], walking /, /home, /var/www, /usr, and /var/tmp. The actors also stage exfiltration data under innocuous file names to minimize detection of MySQL email dumps, including 001.gif, Css.js, Include.png, M2k.js, M2k_list.js, and M2k_ui_adm.js. Exfiltration The PHP bot automatically uploads collected emails to a remote server [ T1020 ], obfuscating the directories and files it creates and renaming its child process to crypto. The main C2 domain for the bot was natcloudservice[.]com. Integrity Technology Group maintains a custom web application that provides third-party access to stolen email content, where users pass arguments in URLs to view mail for specific accounts. Exploited Vulnerabilities Appendix B of the advisory lists eight CVEs successfully exploited by Integrity Technology Group, recovered from the MicroScan penetration testing scripts. Five of these were newly added to CISA’s Known Exploited Vulnerabilities (KEV) Catalog with this advisory, marked below with an asterisk: CVE-2014-6278 (GNU Bash, remote code execution) CVE-2015-3306 * (ProFTPD 1.3.5, unauthorized read) CVE-2015-5477 * (ISC BIND 9.x, denial of service) CVE-2016-3081 * (Apache Struts, remote code execution) CVE-2019-11510 (Pulse Connect Secure, unauthorized read) CVE-2021-22205 (GitLab 11.9 and later, remote code execution) CVE-2021-3199 * (ONLYOFFICE DocumentServer 5.1.5 through 5.6.2, unauthorized write) CVE-2023-22894 * (Strapi up to 4.5.5, information disclosure) The age of this list is itself the finding. Several of these vulnerabilities are a decade old, and their presence in a 2026 advisory means the actors are still successfully exploiting them against internet-facing systems that were never patched or never replaced. Integrity Technology Group IOCs The advisory provides extensive IOCs across Appendix A, including: Observed domain names and IP addresses associated with the threat activity Domain names attributed to threat activity and obfuscation network hosts, annotated by type (infrastructure, SoftEther host) with first-seen and last-seen dates Leveraged webshells, binaries, and scripts Potentially dropped files For the full list of IOCs, download the AA26-281A STIX XML or STIX JSON bundles referenced in the advisory. The authoring organizations recommend vetting older indicators before blocking—some date back to 2016, and their current ownership may have changed. Organizations should hunt for the tooling and artifacts named throughout the advisory, including SoftEther installers renamed to conhost.exe or dllhost.exe, processes named DiagTrack.exe making outbound HTTP connections, DC.exe, office-cli, and the staging file names used for MySQL email dumps. SafeBreach Coverage & Playbook Attack Updates Existing Behavioral Coverage SafeBreach already simulates several of the attacker behaviors detailed in this campaign. The following 15 existing attacks are now mapped to Integrity Technology Group: Recon / Web-App Exploitation #8019: Port scanning target simulator #8018: Web application port scanning #6604: XSS- Cross Site Scripting via URL parameters #6605: SQLi- SQL injection via URL parameters #6877: Path Traversal Attack- GET request #4271: Remote exploitation of Pulse VPN vulnerability CVE-2019-11510 (WAF) Credential access #5833: Extract Login Information using MimiKatz DCSync #6801: Credential harvesting using Mimikatz DCSync with interactive session token #6802: Credential harvesting using Mimikatz DCSync with user credentials #7169: Dump the SAM database from the registry (Windows) #9456: Extract SAM credentials from registry #258: Brute force attack over the HTTP/S protocol Defense evasion #2248: Masquerading Collection / Exfiltration #6975: Covert archived data asset exfiltration using HTTP/S POST #107: Covert data asset exfiltration using HTTP/S POST These simulations are mapped to MITRE ATT&CK and can be used immediately to validate detection and response capabilities across your infrastructure. New IOC-Based Coverage To address the newly released IOC guidance, SafeBreach has added the following 11 new attacks: #11835: Email bx_shell (38ade9) webshell as a compressed attachment #11825: Write fscan (2480e9) hacktool to disk #11826: Pre-execution phase of fscan (2480e9) hacktool (Windows) #11827: Transfer of fscan (2480e9) hacktool over HTTP/S #11828: Transfer of fscan (2480e9) hacktool over HTTP/S #11829: Email fscan (2480e9) hacktool as a compressed attachment #11830: Email fscan (2480e9) hacktool as a compressed attachment #11831: Write bx_shell (38ade9) webshell to disk #11832: Transfer of bx_shell (38ade9) webshell over HTTP/S #11833: Transfer of bx_shell (38ade9) webshell over HTTP/S #11834: Email bx_shell (38ade9) webshell as a compressed attachment How to Test Your Defenses Against Integrity Technology Group with SafeBreach SafeBreach customers can now validate their security controls against these TTPs in multiple ways. Method 1 – SafeBreach Scenarios: Navigate to the SafeBreach Scenarios page and choose CYBERSECURITY ADVISORY Integrity Tech. Method 2 – Attack Playbook: From the Attack Playbook, filter by the threat name CYBERSECURITY ADVISORY Integrity Tech or search for the attack IDs listed above. Method 3 — Known Threats Series: Open the Known Threats Series report, select the CYBERSECURITY ADVISORY Integrity Tech report, and run the simulations. Additional Advisory Steps What are mitigation strategies for Integrity Technology Group? The authoring organizations recommend the following, prioritized here for the behaviors in this advisory: Require multifactor authentication (MFA) across all services to the extent possible, particularly for webmail, VPNs, and accounts that access critical systems. Patch and update software and firmware promptly, prioritizing the eight CVEs in Appendix B, and replace end-of-life products with supported alternatives. Sanitize user input in web applications to prevent XSS payload injection, and review web application access logs for directory traversal, command injection, and enumeration attempts. Disable unused services and ports, and configure applications to reveal as little information as possible in login pages and banners. Replace default passwords, limit and audit accounts with administrative privileges, and apply least privilege. Implement network segmentation so a compromised device cannot reach another organizational unit’s sensitive resources. What are proactive threat monitoring activities for Integrity Technology Group? Hunt for unauthorized SoftEther installations. Look for SoftEther clients configured to reconnect on startup, especially installers named conhost.exe or dllhost.exe in non-standard paths. Watch for unexpected Active Directory replication. DCSync activity from any host that is not a domain controller is the single highest-value detection for this campaign. Instrument every Exchange authentication interface. Password spraying through OAB, EWS, MAPI, Autodiscover, and ActiveSync frequently bypasses monitoring that only covers OWA. Flag masqueraded processes. Alert on known Windows binary names running from unexpected paths or making unexpected outbound HTTP connections. Monitor for high outbound upload volumes from workstations and servers that normally upload far less than they download—the signature of staged mailbox exfiltration. Stay Ahead with SafeBreach For a complete view of your security gaps against Integrity Technology Group’s attacks, sign into SafeBreach and run the latest simulations mapped to CYBERSECURITY ADVISORY Integrity Tech. In addition to testing your defenses against the reconnaissance and web application exploitation followed by credential access, masquerading, and covert exfiltration outlined in this advisory, you can go a step further with SafeBreach Propagate . Propagate enables you to assess how attackers could pivot across your environment post-compromise. It maps high-risk attack paths, visualizes lateral movement, and prioritizes remediation efforts based on exposure to your most critical assets. Learn more in the SafeBreach Propagate solution brief . About the Author Noam Sagiv is a Product Manager at SafeBreach leading attack content for the SafeBreach CTEM Platform. Before SafeBreach, she spent five years in Israel’s Unit 8200, where she led teams building data-driven intelligence products used by more than 10,000 users. Frequently Asked Questions What is Joint Cybersecurity Advisory AA26-281A? AA26-281A is a joint Cybersecurity Advisory published October 8, 2026, and led by the FBI, co-sealed by CISA, the NSA, and cybersecurity agencies in the United Kingdom, Australia, Canada, Japan, New Zealand, and Spain. The advisory describes how Chinese government-linked threat actors , enabled by the China-based Integrity Technology Group, combine automated scanning, botnets, and hands-on exploitation to steal sensitive data. Targets include US critical infrastructure, law enforcement, education, and religious organizations, as well as victims across Southeast Asia, Africa, and North America. What tools and infrastructure do the AA26-281A threat actors use? The AA26-281A threat actors rely almost entirely on open source and legitimate software. For reconnaissance they use BBScan, dirsearch, Fscan, ksubdomain, masscan, Nmap, OneForAll, ShuiZe, and WPScan, plus a custom platform called MicroScan containing over 1,300 penetration testing scripts. They use EBurst against Microsoft Exchange, SoftEther VPN for persistence, DC.exe for DCSync, and a PHP bot and the office-cli utility to collect and exfiltrate email. Which vulnerabilities are being exploited in AA26-281A? Appendix B of AA26-281A lists eight successfully exploited CVEs: CVE-2014-6278, CVE-2015-3306, CVE-2015-5477, CVE-2016-3081, CVE-2019-11510, CVE-2021-22205, CVE-2021-3199, and CVE-2023-22894. Five of these—CVE-2015-3306, CVE-2015-5477, CVE-2016-3081, CVE-2021-3199, and CVE-2023-22894—were newly added to CISA’s Known Exploited Vulnerabilities Catalog alongside the advisory. All eight affect internet-facing products, and several are more than a decade old. Who is behind the attacks outlined in AA26-281A, and what are they doing? The activity is attributed to Chinese government-linked threat actors enabled by the Integrity Technology Group, a China-based for-profit company with links to the Chinese government. The authoring organizations state the actors use TTPs consistent with activity publicly tracked as Flax Typhoon, Ethereal Panda, and Red Juliett. Their objective is data theft, primarily email: they collect account credentials and exfiltrate mailbox data from on-premise systems and cloud services, in some cases restricting access to stolen data to IP addresses in Xiamen, China. How can organizations test their exposure to the threats outlined in AA26-281A? SafeBreach customers can validate controls against AA26-281A three ways. Navigate to the SafeBreach Scenarios page and select CYBERSECURITY ADVISORY Integrity Tech; filter the Attack Playbook by CYBERSECURITY ADVISORY Integrity Tech; or open the Known Threats Series report, select the CYBERSECURITY ADVISORY Integrity Tech report, and run the simulations. Each path executes simulations mapped to the advisory’s MITRE ATT&CK techniques so teams can confirm whether their controls detect the reconnaissance and web application exploitation followed by credential access, masquerading, and covert exfiltration outlined in the advisory. You Might Also Be Interested In Research Blog Countering Chinese State-Sponsored Espionage Campaigns: SafeBreach Coverage for CISA Advisory AA25-239A Read More Solution Brief SafeBreach CTEM Platform Read More Blog Volt Typhoon: SafeBreach Coverage for US-CERT Alert (AA24-038A) Read More Get the latest research and news

apt检测报告 @safebreach-blog source ↗

Oct 6, 2026 AI Has Cut Post-Compromise Timelines to Minutes. Your Validation Cadence Needs to Catch Up. The Microsoft Digital Defense Report 2026 shows why continuous threat exposure management (CTEM), not point-in-time testing, is now the necessary standard to prove your security controls work. Summary Microsoft’s Digital Defense Report 2026 warns that AI can shorten post-compromise activity—secret discovery, lateral movement, data exfiltration—from days to minutes. Phishing jumped from 7% to 23% of initial access in Microsoft’s incident response engagements year-over-year; exploitation of public-facing applications climbed from 15% to 24%. Government organizations absorbed 27% of all tracked attacks, with the US alone accounting for 25.5% of global attack volume. If attackers can go from foothold to impact in minutes, a security program that validates its controls once a quarter—or once a year—is validating a version of its environment that no longer exists by the time the report is finished. The fix isn’t more point tools. It’s continuous threat exposure management (CTEM): ongoing, automated validation that tells you in real time whether your actual attack paths are closed, not whether they were closed last time someone checked. Author: Michael De Groat, Vice President of Sales Engineering For years, security leaders have operated on a reasonable assumption that attackers need time: time to gain a foothold, time to escalate privileges, time to move laterally, and time to find what they’re after. That assumption shaped how most organizations validate their defenses—quarterly reviews, annual penetration tests, occasional red team engagements—because the risk window between “we tested this” and “an attacker is exploiting this” felt manageable. The newly released Microsoft Digital Defense Report 2026 says that assumption no longer holds. According to the report, threat actors are now using AI to compress what used to be a days-long post-compromise sequence—credential discovery, lateral movement, data exfiltration—into a matter of minutes. The strategic advantage of speed that attackers have always chased just got a force multiplier, and most security programs’ testing cadence hasn’t moved to match it. Key Findings from the Microsoft Digital Defense Report 2026 A few numbers from the report describe where the pressure is actually building: Phishing as an initial access vector rose from 7% to 23% of initial access in Microsoft’s incident response engagements—a more than threefold increase, and a reminder that identity remains the easiest door in. Exploitation of public-facing applications climbed from 15% to 24% , continuing a trend we’ve seen dominating headlines. Government organizations absorbed 27% of all tracked attacks , the single most-targeted sector, with the US representing 25.5% of total global attack volume. Key Shifts in Microsoft’s 2026 Digital Defense Report Initial Access Vector 2025 2026 Change Exploitation of public-facing applications 15% 24% +9 pts Phishing 7% 23% +16 pts (more than 3x) Valid accounts 17% 13% -4 pts Social engineering 15% 7% -8 pts * Source: Microsoft Digital Defense Report 2026, p. 33. Reporting periods: July 2024–June 2025 and July 2025–June 2026. These findings describe a broadening, accelerating attack surface, hit by an adversary that can now work at machine speed once it’s inside. Vulnerability Management vs. Exposure Management The key takeaway here for most security programs: most of the controls exposed by faster attacks aren’t missing from your stack. They’re simply untested in the way that matters. Most organizations have an endpoint detection and response (EDR) solution, multi-factor authentication (MFA), and segmentation policies (at least on paper). What they often don’t have is current, evidence-based proof that those controls actually stop the specific attack paths adversaries are using right now—this week, not last year’s red team scope. That’s the difference between vulnerability management and exposure management, and it’s exactly the shift organizations are making as they look to move continuous threat exposure management (CTEM) from a theoretical framework to a disciplined practice. CTEM isn’t a rebrand of vulnerability scanning. Continuous threat exposure management is a standing, automated discipline that simulates real attack behavior against your actual environment, continuously, so you have a current answer about how your controls are working, instead of a quarterly one. When attacker timelines were measured in days, a monthly or quarterly validation cycle created some margin for error. Now that it’s being measured in minutes, that margin is effectively gone. The only way to close that gap is to validate continuously, at machine speed, against the same techniques attackers are actually using. What This Means for Your Program Today A few concrete moves that are worth consideration in light of this report: Treat identity controls as a continuous validation target, not an annual one. Microsoft’s report calls for passkeys, phishing-resistant MFA, and reduced privilege—but a policy on paper isn’t a tested control. Validate that your identity controls hold up against current phishing and credential-theft techniques on an ongoing basis. Assume your public-facing applications are being probed right now. With exploitation of these systems now the top initial access vector in Microsoft’s incident response engagements at 24%, “we’ll test the perimeter next quarter” is not a defensible posture. Validate the full attack path, not just the entry point. AI-accelerated post-compromise movement means the lateral steps after initial access matter as much as the initial access itself. Attack path validation —mapping and testing what an attacker can actually reach once they’re in, not just how they got in—is what turns “we have segmentation” into “we’ve proven our segmentation holds.” The Bottom Line Speed is now the attacker’s advantage and the only realistic defense is a validation cadence that matches it: continuous rather than periodic, evidence-based rather than assumed, and run against the real techniques adversaries are using today. That’s the premise the SafeBreach CTEM Platform is built on, and it’s exactly what our breach and attack simulation (BAS) and automated attack path validation capabilities are designed to deliver—proof, on an ongoing basis, that when an attacker moves in minutes, your controls will respond as expected. If the last full validation of your attack surface happened more than a quarter ago, it’s worth asking what’s changed since then—because according to Microsoft, the adversary may only need a few minutes to find out first. Ready to see where your environment actually stands against today’s fastest-moving attack techniques? Schedule a demo to see how the SafeBreach CTEM platform validates your real-world risk, continuously, so you’re never operating on last quarter’s answer about where your defenses stand. About the Author Michael De Groat is the VP of Global Sales Engineering at SafeBreach, where he helps enterprises turn adversarial exposure validation (AEV) into a clear, continuous view of their security posture. He joined SafeBreach in 2016 as the company’s first sales engineer and built the function from the ground up over the last decade. Before SafeBreach, Michael held sales and solutions engineering roles at Trend Micro, IBM, Netskope and Imperva. Earlier in his career he worked in technical roles at Lockheed Martin and Elevations Credit Union. He holds a degree in Computer Information Systems from Colorado State University. Frequently Asked Questions What is continuous threat exposure management (CTEM)? Continuous threat exposure management is an ongoing, automated discipline of testing your actual environment against real attack techniques—rather than relying on periodic scans or annual penetration tests—so you have a current, evidence-based answer to whether your security controls actually hold up against today’s threats. How much has AI sped up cyberattacks? According to Microsoft’s Digital Defense Report 2026, AI can shorten post-compromise activity—secret discovery, lateral movement, and data exfiltration—from a timeline of days down to minutes. What were the biggest initial access vectors in Microsoft’s 2026 report? According to Microsoft’s Digital Defense Report 2026, phishing rose from 7% to 23% of initial access in Microsoft’s incident response engagements, and exploitation of public-facing applications climbed from 15% to 24%—making identity and perimeter applications the two fastest-growing entry points. Which sector was targeted most in the report? Government organizations absorbed 27% of all tracked attacks, the most of any sector, with the United States accounting for 25.5% of total global attack volume. Why isn’t an annual penetration test enough anymore? A penetration test is a snapshot of one point in time. If attackers can move from initial foothold to full impact in minutes, the environment a pentest validated can look meaningfully different by the time that report is even finished—let alone a quarter or a year later. What’s the difference between vulnerability management and exposure management? Vulnerability management asks whether a known CVE has been patched. Exposure management asks a broader question: given everything in your environment right now, which attack paths are actually reachable and exploitable—regardless of whether a CVE exists for them. How does attack path validation fit into CTEM? Attack path validation tests what an attacker could actually reach after gaining a foothold—lateral movement, privilege escalation, access to critical assets—so you’re proving the full chain is blocked, not just the entry point. How often should organizations validate their security controls? Given how fast AI-assisted attacks now move, validation needs to be continuous rather than periodic—run frequently enough that the answer to “are we exposed” reflects your environment as of today, not last quarter. You Might Also Be Interested In Blog AI-Powered Ransomware Attacks Are Here: What CISOs Need to Validate Now Read More Blog What the OpenAI-Hugging Face Incident Reveals for Every Security Program Read More Blog Have Frontier AI Models like Claude Mythos and GPT-5.5 Really Changed the Zero-Day Lifecycle? Read More Get the latest research and news

llm报告事件 @safebreach-blog source ↗

Threat Adaptation: Your Future SOC Metric By: Brian Moran, VP of Product Marketing October 13, 2026 For years your security operations center kept score with two core numbers: mean time to detect and mean time to respond. AI-powered threats require a more proactive and preemptive metric: “threat adaptation.” A recent Gartner® report introduced threat adaptation as a new metric in its guidance toward building the SOC of the future. Threat adaptation (TA): The time it takes from when your organization is aware of a new threat until the environment is assessed, validated (including defensive posture) and an automated stopgap adaptation is made (such as implementing new detection code, limiting access, throttling, etc). This measure should be taken at the most advantageous position (such as defensive team, exposure team, or identity and access team) and seen as a unified collaborative response. Gartner, The Security Operations Roadmap to 2030, Kevin Schmidt, Alex Tytarenko, Steve Santos, Eric Ahlm, Pete Shoard, October 4, 2026 Why Existing Metrics Fail In my time running product management for managed detection and response, I’ve seen the value of mean time to detect (MTTD) and mean time to respond (MTTR) in measuring SOC performance on its original core duty: reactive security. But both measure how quickly a team handles an attack once it reaches the environment. Neither tells security leaders how quickly their organization can prepare for a newly discovered threat, test its defenses and close the gaps before an attacker exploits them. AI increases both the scale and speed of new threats. We’ve seen the evidence and power of frontier AI to: Discover new exposures Turn vulnerabilities into usable attack techniques Breakout from an initial foothold to achieve attack objectives Coordinated agents that work in parallel to achieve attack objectives Faster detection and response alone leave an important question unanswered: How long are we exposed after we learn about a threat? The Building Blocks of Threat Adaptation Know your attack surface . Maintain a current view of assets and exposures as your environment changes. When a new threat emerges, this context helps teams identify which systems may be affected and where to focus testing. Threat awareness . Track active campaigns and emerging attacker techniques to understand which threats are relevant to your organization. Translate that intelligence into specific attack behaviors your teams can test against the environment. Security controls . Maintain visibility into the configuration and changes of the controls that protect your environment. This gives teams the context to assess existing defenses and identify where prevention or detection may need to be updated. Validate threats and controls . Use automated attack simulations to prove whether security controls can prevent and detect the attacks associated with each exposure. Test the SOC response as well to establish whether teams can investigate and contain the activity that gets through. Virtual patch . Apply threat-specific prevention updates to security controls to reduce exposure while a permanent fix is prepared. These stopgap measures can block an attack technique or restrict the access it needs to succeed. Detection engineering . Build, test and deploy detection logic for relevant attacker behaviors that existing rules miss. Validate that the new logic produces actionable alerts with enough context for the SOC to investigate and respond. Validate updated defenses . Re-test the environment after prevention and detection changes to confirm they address the demonstrated gaps. Use the results to verify effective mitigation and identify any remaining work. Adapt at Machine Speed with Agentic Cyber Defense Engineering Many mature security operations teams already have these building blocks in place. The challenge is to make them work together in an automated and near-autonomous process that adapts security controls to new threats at machine speed. At Cymulate, we call this agentic cyber defense engineering – a closed-loop system to prove, prioritize and adapt. Agentic cyber defense engineering optimizes your existing security to stay ahead of the threats targeting your environment and shorten your threat adaptation time. This approach applies AI to execute repeatable and continuous routines that test threats, validate exposures, push control updates for immediate prevention and build new detection logic – all to engineer defenses that adapt to current threats. With a 10-year history of advanced offensive security, the Cymulate platform delivers agentic cyber defense engineering that: Profiles relevant attacker behavior Tailors testing to your environment, assets, industry, exposures and controls Executes safe, realistic attack simulations Validates prevention and detection of existing security controls Prioritizes fixes by real-world impact Optimizes security controls with prevention and detection updates that mitigate exposures and threats Agentic Cyber Defense Engineering at a Glance Requirement Required Inputs Expected Outcome 1. Profile — Integrate and maintain a living contextual model Asset discovery, exposure management, threat intelligence, security control and telemetry feeds A continuously current model of assets, exposures, architecture, controls and relevant attacker behavior that automatically advances the closed loop 2. Tailor — Configure organization-specific assessments Maintain current environment/threat profile and configure test points, environments and tool integrations An attack assessment scoped and configured to the organization's specific environment, without manual setup 3. Execute — Safely run realistic attacks Tailored attack scenarios, defined scope and safety boundaries, approval and escalation policies Controlled, repeatable execution of realistic attacks with no disruption to business operations 4. Validate — Measure prevention and detection Attack execution results with integrated security control and telemetry data Measurable evidence of what was prevented, detected or missed 5. Prioritize — Analyze gaps by demonstrated risk True risk scores by correlating validation evidence, threat intelligence, exploitability, attack-path reachability and business impact data A stacked ranked list of exposures based on demonstrated impact rather than static severity scores 6. Optimize — Configure controls and improve security Mitigate priority exposures with vendor specific prevention and detection improvements Auto-deployed, governed mitigations with automatic re-validation to ensure the mitigation was successful Learn More about How Cymulate Drives Threat Adaptation To plan a future SOC built for threat adaptation, read the Cymulate e-book Agentic Cyber Defense Engineering . Brian Moran, VP of Product Marketing With over a decade in cybersecurity product marketing managerial positions, VP of Product Marketing Brian Moran is a driven product manager and product marketer with a track record of launching new innovative products and reigniting the growth of mature portfolios. More about Author Table of Contents Why Existing Metrics Fail The Building Blocks of Threat Adaptation Adapt at Machine Speed with Agentic Cyber Defense Engineering Agentic Cyber Defense Engineering at a Glance Learn More about How Cymulate Drives Threat Adaptation “Cymulate enables us to test Nemours’ defenses against the latest cyber threats as they emerge, prioritize remediation efforts, and improve our security team’s incident response skills.” Jim Loveless , CISO Nemours Learn More Featured Resources View More Resources Webinar From AI Hype to AI Risk: A New Security Reality Cymulate CTO Avihai Ben-Yossef explains AI-driven attacks and how continuous validation helps defenders prioritize. Watch Now blog Inside the Agentic Cyber Defense Engineering Architecture: How it Works Understand the environment, coordinate specialized security work and act across the security stack. Read More Page Cymulate for Security Operations Validate threats and understand gaps to build stronger defenses that protect your biggest sources of threat exposure.​ Read More GET A PERSONALIZED DEMO Ready to see Cymulate in action? Book a Demo

议题报告 @cymulate-blog source ↗

PSIRT Double free in signal handler Summary A Double Free vulnerability [CWE-415] in FortiOS, FortiProxy and FortiSwitch may allow a local authenticated attacker to crash the newcli daemon via crafted CLI commands. Version Affected Solution FortiOS 8.0 Not affected Not Applicable FortiOS 7.6 7.6.0 through 7.6.7 Upgrade to upcoming 7.6.9 or above FortiOS 7.4 7.4 all versions Migrate to a fixed release FortiOS 7.2 7.2 all versions Migrate to a fixed release FortiProxy 7.6 7.6.0 through 7.6.6 Upgrade to 7.6.7 or above FortiProxy 7.4 7.4 all versions Migrate to a fixed release FortiProxy 7.2 7.2 all versions Migrate to a fixed release FortiSwitch 8.0 Not affected Not Applicable FortiSwitch 7.6 7.6.0 through 7.6.6 Upgrade to 7.6.7 or above FortiSwitch 7.4 7.4 all versions Migrate to a fixed release FortiSwitch 7.2 7.2 all versions Migrate to a fixed release Follow the recommended upgrade path using our tool at: https://docs.fortinet.com/upgrade-tool Acknowledgement Fortinet is pleased to thank The UK's National Cyber Security Centre (NCSC) for reporting this vulnerability under responsible disclosure. Virtual Patch No Timeline 2026-10-13: Initial publication IR Number FG-IR-26-182 Published Date Oct 13, 2026 Component CLI Severity Low Discovered External Attack Type Authenticated Known Exploited No CVSSv3 Score 2.1 Impact Denial of service CVE ID CVE-2026-106175 Download CVRF CSAF

dos固件 @forti-ir source ↗

PSIRT HTML injection in WebUI Summary An improper neutralization of script-related HTML tags in a web page in FortiManager and FortiAnalyzer may allow an attacker to perform a cross site scripting (XSS) attack via the Identify Provider name field. Affected Products FortiManager version 6.2.0, 6.2.1, 6.2.2, 6.2.3, 6.2.4, 6.2.5 and 6.2.6 FortiAnalyzer version 6.2.0, 6.2.1, 6.2.2, 6.2.3, 6.2.4, 6.2.5 and 6.2.6 Solutions Please upgrade to FortiManager 6.4.0 or above Please upgrade to FortiAnalyzer 6.4.0 or above Acknowledgement Fortinet is pleased to thank Danilo Costa from PBI Dynamic IT Security for reporting this vulnerability under responsible disclosure. Virtual Patch No Timeline 2026-10-13: Initial publication IR Number FG-IR-20-005 Published Date Oct 13, 2026 Component GUI Severity Medium Discovered External Attack Type Unauthenticated Known Exploited No CVSSv3 Score 4.9 CVSSv3 Temporal Score 4.9 Impact Execute unauthorized code or commands CVE ID CVE-2020-12811 Download CVRF

xss固件 @forti-ir source ↗

PSIRT Persistent Denial of Service in Admin Profile Summary An Uncaught Exception vulnerability [CWE-248] in FortiSandbox may allow a privileged attacker to deny access to the GUI and CLI persistently via crafted HTTP requests. Version Affected Solution FortiSandbox 5.2 5.2.0 through 5.2.1 Upgrade to 5.2.2 or above FortiSandbox 5.0 5.0.0 through 5.0.6 Upgrade to 5.0.7 or above FortiSandbox 4.4 4.4 all versions Migrate to a fixed release Acknowledgement Fortinet is pleased to thank Jason McFadyen of TrendAI Research for reporting this vulnerability under responsible disclosure. Virtual Patch No Timeline 2026-10-13: Initial publication IR Number FG-IR-26-176 Published Date Oct 13, 2026 Component GUI Severity Medium Discovered External Attack Type Authenticated Known Exploited No CVSSv3 Score 4.8 Impact Denial of service CVE ID CVE-2026-106170 Download CVRF CSAF

doscve @forti-ir source ↗

推理与性能 AI 产品与平台 Anthropic 在 官方成本指南 里面提供了一个选择模型的策略:不要用每token的价格来比较模型,而应该看每个任务被解出来的成本是多少。从这个角度出发,你会发现价格表上贵的模型其实往往更便宜。例如在软件工程的评测集SWE-bench Pro上,把Fable 5.1每次出力大小的 effort 设为 low,可以解出88.6%的题目,每个题目花费了0.54美元。相比之下,如果用默认设置的Sonnet 5,只能解出来77.4%的题目,但每解出一个题目需要0.84美元。虽然单价差了大约5倍,但总得来说贵的模型更便宜。 很多团队一旦看到账单高了,第一反应是把某些步骤换成单价更低的模型。这个直觉默认了成本来自单价。但真实 agent 负载里这个假设站不住。更强大的模型往往用更少的轮次、更少的工具调用、更少的重试就做了同样的事,省下来的工作量直接把单价的差价抹平了。 你付的是任务,不是 token 关于推理的成本,大家通常很直观地会去看价格表上单价是多少。但当你让 agent 去修一个 bug、出一份报表的时候,你购买的是确定的结果,而不是按字收费的字符。agent 的工作量和普通的文本补全接口完全是两码事,账单应该按结果算,不应该按单价算。 Anthropic 关于成本和智能的优化文档里有一组数据说明了这一点。按每解出一个任务来看的话,又是一副图景。默认配置下,Opus 5.5 可以解出 92.8% 的任务,每解出一个任务只需要 0.22 美元。Fable 5.1 也可以解出 92.3% 的任务,但每解出一个任务需要 1.19 美元。注意 Fable 5.1 每 token 的价格明显更高($10 / $50 对比 Opus 5.5 的 $4 / $20)。但这里在准确率这个噪声范围内,更贵的模型并没有带来更好的结果,反而每任务成本高了约五倍。 模型的迭代也遵循同样的规律。根据文档中的测试,从 Sonnet 4.6 切换到 Sonnet 5,解题率提升了 5 个百分点,每解出任务的平均成本还降低了 15%;从 Fable 5 升级到 Fable 5.1,解题表现相当,每解出任务的花费降了 43%。 但换更强模型不等于账单自动下降。在文档的另一个基准 DeepResearch Bench II 上,Fable 5.1 把 effort 调到 high,准确率虽比前代高 2~3 个百分点,每解出任务成本却涨了 41%。单任务成本能否降低,不取决于模型有多强,而取决于这类任务中模型做的工作量能否被收住。 价格表上的 token 费率只作用在被打折的输入上,最终账单由模型的活动量决定 贵单价乘在一个打折的数上 贵模型能在每任务成本上赢便宜模型,是因为 agent 的运行方式把费率的作用架空了。agent 每跑一轮就攒一层状态,下一轮又要把整段对话历史、工具定义、系统提示词重新发给模型。一个跑 40 轮的任务,等于把第一轮的内容发了 40 遍。不做优化的话,输入 token 会随轮数平方增长。 在生产环境里,提示词缓存承担了绝大多数重复内容。Anthropic 公布过一组来自其第一方 API 的数据:对连续 14 天里 10 万个以上的组织进行观察,在 agent 循环里,从提示词缓存里读的 token 占所有输入 token 的中位数是 84.2%,上四分位是 91.7%。在写代码的组织里中位数是 88.5%,当单任务工具调用超过 25 次时升到 90.9%,头部十分之一的组织更超过 95.9%。任务一旦跑到深处,好的 agent 循环里只有不到 1% 的输入 token 需要付全价。 但缓存读便宜很多。普通模型的缓存读一般只要输入价格的十分之一,Opus 5.5 和 Sonnet 5.5 是二十分之一,Fable 5.1 甚至只有四十分之一。因为缓存默认只存 5 分钟,而 agent 的轮次往往只是几秒,所以每轮基本都能吃到这个折扣。因此,价格表上那个吓人的输入单价,在实际中大多数时候乘的是一个打折到一折甚至零点二五折的小数。真正决定账单的是模型跑了多少轮、调了多少次工具,而不是标价。 所以在会话中间换模型来省钱,往往适得其反。因为提示词缓存要求前缀逐字节完全相同,你在中间换了一个更便宜的模型,或者改变了系统设定,整条缓存就会直接作废。 Anthropic 做过实验:10 万 token 的前缀断一次,就要按写入价重算全量。在 Fable 5.1 上这一刀要花 1.25 美元,原本读一次只要 0.03 美元,差了 50 倍;Opus 5.5 和 Sonnet 5.5 上也差了 25 倍。为了省某一步那点费率去切模型,换来的是几十倍重写代价的缓存断裂。 相比之下,调模型的工作量是顺手得多的杠杆。Anthropic 的 effort 文档 给了五档:low、medium、high、xhigh、max,控制模型在思考、写文本、调工具上花多少力。更关键的是,开一个请求头就能在会话中途改 effort,而且完全不会破坏缓存。 二者结合,直觉就很明显了:想省钱,把同一个模型的 effort 调低,比换一个更便宜的模型划算很多,因为换模型会引发缓存断裂,而调 effort 不会。 effort 曲线不是一根线,是两根 那如果在一个模型上调节 effort 又省事又不破坏缓存,是不是都应该把 effort 设到最小呢?不是的。实测里 effort 的效果在不同的任务上有两种完全不同的表现。 第一类是研究和知识抽取的活。Anthropic 在 WideSearch、DeepWideSearch、BrowseComp、GDPval 几个检索分析基准上测试了 Fable 5,这些基准官方默认档是 high,往下调 effort 准确率几乎不动,成本却能一路降下来。 对于这种研究和知识抽取的任务,把 effort 从 high 调到 low,准确率只掉了 1 到 3 个百分点,但每解出一个任务的成本可以省掉三分之一到一半。中档(medium)的准确率和默认几乎一样,成本只有默认的 70% 到 87%。在四个基准里,默认的 high 相比 medium 都没有可测的精度收益。降 effort 还更快,DeepWideSearch 上每题从默认的 7.9 分钟缩到 4.5 分钟。在这些活上硬撑高 effort,纯属浪费算力。 第二类是长程工程和复杂编码,曲线很陡。在 Opus 5.5 解 SWE-bench Pro 代码缺陷的测试中,这个模型默认档是 medium,准确率高度依赖给它多少思考和计算量。medium 比 high 低约 2.5 个百分点,成本是 high 的 70%;降到 low,解题率掉约 8 个百分点,从 high 的 95.3% 掉到 87.4%,成本降到约三分之一,也就是 0.11 对 0.29 美元;升到 xhigh,准确率加约 1.4 个百分点达到 96.7%,但成本涨到 0.74 美元,是 high 的 2.5 倍。 研究类任务的 effort 曲线近乎平缓,而长程编码任务的 effort 曲线表现陡峭 把一档 effort 死磕在一个模型上,研究任务上浪费预算,长程编码上压太低又大面积失败。 对于这种陡峭型任务,如果有明确的验证信号(单测过不过/编不编得过),一个聪明的办法是:全部先用low跑,失败了的再用high重跑一遍。SWE-bench Pro上,全量high是95.3% / 0.29美元/题;先low再重跑失败题,准确率反而到了97.0%,每题只要0.17美元。代价是失败题要跑两遍,好处是用便宜算力过掉大部分常规题,把高effort只留给真正难的分支,总成本省了近一半,准确率还略高。 那多模型分层什么时候还有优势 那既然用一个模型内部调整 effort 就能实现类似缓存和算力调节的效果,那为什么还要用多模型分层架构呢?Anthropic 在文档里有一个硬性标准:一个多模型分层架构要想成立,得先证明它比用同一个模型降低 effort 更划算。有很多看起来精巧的多模型分层方案,做简单的数学比较后发现,其实比用同一个模型降低 effort 还贵。 举个例子,比如顾问模式。用一个轻量级的执行者跑主循环,卡了再请更聪明的高阶模型当顾问指点。要见效,需要二者能力差距够大,且执行者真会去问。 但问题在于,执行者到底会不会问?这个恰恰是最薄弱的一环。在GPQA Diamond的198题里,官方用单跑85%的Haiku 5.5和单跑91%的Sonnet 5.5当执行者,Opus 5.5当顾问。两轮一共396次机会,执行者一次没问(0/396)。只是挂了顾问工具的定义,每次请求就多约1000个输入token,把Haiku每题成本涨12%,Sonnet涨25%,得分却和执行者单跑一样。 更麻烦的是,一旦把执行者调到低 effort 档位试图省成本,它连自己卡住了都常察觉不到,问得更少,表现甚至不如单跑。在另一项 370 题的内部代码基准上,官方测了 Opus 5.5 当执行者、更前沿档位的 Fable 5.1 当顾问:比 Opus 5.5 单跑 high effort 多 1.7 个百分点,每次尝试成本却是 2.1 倍(2.92 美元)。这点提升跟直接把 Opus 5.5 开到 xhigh 差不多,多出来的架构没换来更好的性价比。 另一个思路叫协调者模式:用最强的模型来做规划和分块,然后把大量细活丢给便宜的 worker。但在常规任务上这样去拆同样不划算。在 DeepWideSearch 上,让 Fable 5 在 low effort 下自己跑,精度和它指挥 Sonnet 5 干的复合架构完全一样,但便宜 29%。原文结论是:单纯把 effort 调低就吊打了整套架构重构。 根据官方测试,协调者模式只有在两种情况下能够真正节省成本: 一是在日常成本的长尾上买保险。很多开放性的搜索或排查任务,强模型会偶尔在某道常规题上打转或过度发挥,极少数这种轮次就把整个周期里的大头账单给吃了。让协调者把常规活甩给便宜模型,就能把最贵的那段长尾给掐掉。在BrowseComp的测试切片里,用Fable 5 + 1个Sonnet 5,平均成本砍了大约一半,90%分位从33美元降到12美元,还把”某个单模型跑了84美元却答错”的极端情况给消掉了。 二是活太大、一个上下文窗口装不下。面对 14 个公开代码包、2160 万 token 的大型缺陷排查,单靠 Fable 5.1 要 15 到 20 小时、468 到 552 美元;换成 1 个 Fable 5.1 带 25 个 Sonnet 5 并行,2.3 小时跑完,总成本降约 50%,代价是得分低 10 到 12 分。 如果你的工作是一条依赖链、能装进一个上下文窗口,或者单模型降 effort 之后已经足够,就别去搭多模型。选型的第一步永远是先给当前模型扫一遍 effort;只有在确实发现能力缺口之后,才给更强的模型在 low effort 下单跑定价,把这个数字当作接下来所有分层方案的门槛。 这套方法的边界,和它没说透的那一半 这套推导可以用,但要记住它的成立条件:数据来自 Anthropic 官方的技术指南,用的是他们做的 benchmark、当时的 list price 和 5 分钟缓存。文档也明确写着这些测量只有方向性,别拿你自己的业务直接套用,要在自己真实流量上重测。 另一个层面是,这篇文章虽然给出技术建议,但它也是厂商自己的立场。Anthropic 在自己的 官方选型指南 里提到 tuning effort 往往比换模型更划算(“Tuning effort is often a better lever than switching models”)。这句话背后有两层。 从工程角度讲,这个费率和活动量、缓存读写、跨模型缓存断裂的分析是有道理的,它确实抓到了agent账单的主要来源,也解释了为什么长轮次上下文里降级模型常常得不偿失。 但从平台方的生意来看,effort 只是厂商在自己产品里定义的一个算力旋钮。不管你在 low 还是 high 上调节,你的计算和流量都在 Anthropic 这里。它推荐你用自家的模型,这样你花钱省、系统跑得好、还不会被带走。真正让厂商亏的事,不是你调一个模型的 effort 从 high 到 low,而是你把整个调度跑到别的模型、甚至开源上面。 那具体怎么落地呢?不是照单全收”不分层、调effort”这个结论。接受账单由工作量决定这个事实,然后在搭建多模型调度、切流量给便宜模型之前,先在你自己的业务里给主力模型画一条effort曲线。把它在low effort下的每任务开销当一条底线;任何分层方案,都得先在真实账本上证明能打赢这条不用付缓存断裂代价的单模型基线。

llmai_agent报告 @yage-share source ↗