⚡Megatron LLM Hub
All digests
Twitter

推特安全流

Top6

恶意或失陷主机即可触发,公开检测规则可直接落地

🪟𝗧𝗲𝗮𝗺𝗩𝗶𝗲𝘄𝗲𝗿 𝗥𝗲𝗺𝗼𝘁𝗲 𝗖𝗼𝗱𝗲 𝗘𝘅𝗲𝗰𝘂𝘁𝗶𝗼𝗻 𝗗𝗲𝘁𝗲𝗰𝘁𝗶𝗼𝗻 ​ On October 29, Remedio security researcher Omri Dar disclosed research into a TeamViewer vulnerability, 𝗖𝗩𝗘-𝟮𝟬𝟮𝟲-𝟭𝟭𝟭𝟱𝟴𝟱. A malicious or compromised host could potentially exploit the vulnerability to crash TeamViewer on the viewer’s system and, depending on the platform, achieve arbitrary code execution. https://t.co/TRz9MHdUUo ​ The KQL query below provides a detection opportunity by identifying instances where 𝘃𝘂𝗹𝗻𝗲𝗿𝗮𝗯𝗹𝗲 𝗧𝗲𝗮𝗺𝗩𝗶𝗲𝘄𝗲𝗿 spawns a shell process. Such activity may be consistent with exploitation and should be investigated alongside other endpoint telemetry. KQL Code: https://t.co/NHrcVG7iy0 ​ #threathunting

rcecve检测 @0x534c source ↗

官方确认影响约 5000 用户,加载即触发,可立即自查扩展版本

Cursor users: check your Claude Code extension. Cursor confirms a trojanized Anthropic.claude-code v2.1.313 in its extension marketplace on 29 October, 09:27 to 20:35 UTC. It estimates up to 5,000 affected users and identifies macOS ARM64 and Windows x64 builds. Loading the extension executes malicious code that collects host information and attempts to install a persistent SSH backdoor. Cursor says SSH access would fail on most hosts due to bugs and insufficient privileges. That limits remote access; it doesn’t undo malicious code execution. Credentials and repositories accessible to that account are potentially at risk. The package has been removed and the C2 taken offline. Uninstalling the extension may still leave persistence behind. If this version ran, investigate the host, remove persistence and rotate accessible credentials from a clean system. Updating the extension alone doesn’t prove the machine is clean. The vendor-confirmed distribution channel is Cursor’s marketplace. Microsoft VS Code uses a separate store. These reports don’t establish a compromised Microsoft Marketplace or malicious npm release. My recommendation: turn off automatic extension updates completely - for yourself and for your developers. With supply chain attacks like this, I don’t think the benefit of another feature or a minor UI fix justifies automatically accepting every new release onto developer machines. Update manually when something stops working, when a bug is actually getting in your way, or when a relevant security fix needs to be applied. Otherwise, leave a working extension alone. Cursor advisory and cleanup instructions: https://t.co/dk9he5dYbG Technical analysis: https://t.co/clSK99WoGz

ai_agent供应链后门事件 @cyb3rops source ↗

把 AI agent 明确列为攻击面,并附 37 个在野漏洞案例可对照。

QiAnXin 2026 Mid-Year Cybersecurity Threat Report is released. APT groups double down on supply chains & edge devices, ransomware goes encryption-less, and AI agents emerge as a new attack surface. 199 reports, 84 actors, 37 in-the-wild vulns analyzed. 👇 https://t.co/Hnf3bnWq06

ai_agentapt供应链报告 @RedDrip7 source ↗

一手事件报告,可核对攻击者究竟拿到多少源码写权限。

@rootxharsh @iamnoooob @HacktronAI Huge kudos to @rootxharsh and @iamnoooob for the research, and to the GNU Savannah maintainers and FSF sysadmins for working with us to get these issues fixed. the FSF has now published its incident report: https://t.co/kv07aUMdKO

供应链报告事件 @S1r1u5_ source ↗
Must-see26

有实时观测页可跟,且已波及多个真实受害者

RT @ImposeCost: Full blown intrusions happening for php2shell right now, which you can watch here: https://t.co/tUVjmvECfM It looks like the actor intends to deploy ransomware. It also looks like they have compromised other actually real victims.

rce在野利用事件 @artem_i_baranov source ↗

若属实,专有漏洞细节与内部数据会流入第三方模型

RT @seanhn: haha, so apparently OpenAI are about to start training on your Daybreak Blue sessions, with no way to opt out. Literally stealing people's private work, as they do it, is cartoonishly evil. It also will presumably kill Daybreak usage by security researchers. Or at least, it should. The thing OAI have had to swear they are not doing to math researchers, they are now explicitly about to start doing to security researchers. Yet another reminder of the importance of open models.

llm数据泄露ai_agent @C2IRIS source ↗

LLM 批量挖洞已能触及主流 OSS,攻防节奏被压缩

It's pretty crazy how easy it is to hunt 0days now.. Two nights of Codex crunching on some leftover credits resulted in three pretty complex disclosures to one of the biggest OSS projects in the world 🥲. Some times we live in

llm漏洞挖掘ai_agent @chvancooten source ↗

跨平台免落盘注入,检测面与利用工具链同时受影响

RT @fridadotre: Frida 17.25.0 is out, and it's a big one. Our out-of-process mapper, born on Darwin, is now cross-platform: agents are mapped straight into the target on Windows and Linux too, without a file on disk or a dlopen() in the target. Plus Efni, Idun, MTE on Darwin, and lots of fixes.

工具注入逃逸windows @0xocdsec source ↗

从漏洞到 root 全链路细节公开,Pixel 6/Android 12 可复现

RT @eshard: We’re publishing a full Android kernel exploit chain built on CVE-2023-20938, a use-after-free in Binder, on a Pixel 6 running Android 12. 👉 Start part 1: https://t.co/u5duFUTXPt https://t.co/BarhiFZI49

提权内核移动端poc @0xocdsec source ↗

未公开漏洞与 PoC 或进训练集,直接影响赏金猎人

RT @whoareme33: OpenAI just told Daybreak Trusted Access users that from Nov 29, Daybreak Blue sessions on personal accounts may be used to train models, even with "improve the model for everyone" turned OFF. If you hunt with it, your unreported bugs, targets, PoCs and techniques become training data. Dupe rate is about to go crazy :)

llmai_agent事件 @pentest_swissky source ↗

沙箱逃逸可组合成完整利用链,Chrome 155 之前版本均处暴露面。

RT @pwn_ai: Introducing CVE-2026-106378: Chrome GPU Sandbox Escape to full Remote Command Execution using synthetic keyboard injection By combining it with common GPU bugs, it becomes a full chain. Binance DevOps was most likely pwnd with a zeroday chain affecting the browser 😬 https://t.co/h87aosIwvm

浏览器逃逸cvepoc @0xocdsec source ↗

已确认首批生产环境实际失陷,利用仍在活跃扫描,需尽快核对版本。

RT @Huntio: PHP2Shell One Week Later: Honeypot Shells on the Web, a Path to Root on Palo Alto Firewalls One week after the #php2shell patch, around 25 internet-facing #PaloAlto firewalls have been rooted, the first confirmed production compromises. On the open internet, the same bug is still only landing shells on honeypots. We pivoted from the public exploitation reports into our AttackCapture open-directory data to see who is running it. What we found: - ~25 internet-facing Palo Alto PAN-OS firewalls rooted, the first confirmed production compromises. Reported by Dream Research, verified by Unit 42, matched in operator logs we captured. - The chain is two bugs: php2shell lands as nobody on the PAN-OS management interface, then a kernel bug (Fragnesia, CVE-2026-46300) escalates to root. - On the open web, watchTowr, 0x4D31, and GreyNoise see shells only on honeypots. Wiz and CrowdStrike found none. - 616 php2shell files across live open directories: a shared ASLR-bypass kit byte-identical across four hosts, a custom heap exploit with a log cleaner, and cloned PoCs. - One operator handed the exploit research to an AI assistant and left the session log exposed. Full research, IOCs, and detection guidance 👇 https://t.co/FUPd8zJX4B

rce事件poc @nas_bench source ↗

Agent 为完成任务会越过从未被告知的边界,护栏设计需覆盖这一层。

Model alignment has a last-mile problem. Post-training instills general norms, but acceptable use is company-specific and rarely reaches the agent's context. An agent optimizing for task completion persists through lines it was never shown. https://t.co/6PsArUw4A5

ai_agentllm报告 @matterpreter source ↗

该库被大量主流运行时与应用打包,且已有活跃利用,影响面覆盖整条软件供应链。

RT @zerodayalpha: ⚡️ 0-Day Alert: Node.js, PyPI, .NET, and others CVE-2026-111497: Google Brotli buffer overflow to code execution The bug is now in active exploitation via popular software that bundles brotli library. https://t.co/DAz7o0044d https://t.co/3X5ePVpXzX

rce供应链cve @cyb3rops source ↗

划定了实际受影响范围,可避免按错误前提排查或过度处置。

There’s misinformation circulating about the Claude Code incident. The confirmed incident involved a tampered copy distributed through Cursor’s own marketplace. A compromise of Anthropic’s official distribution, Microsoft’s VS Code Marketplace or Open VSX has not been established. https://t.co/dk9he5dYbG

ai_agent供应链事件 @cyb3rops source ↗

给编码 Agent 管理员权限的量产式越权案例,权限边界设计可直接参照

RT @choblin29: Holy Shit: WIRED reports an Accenture contractor gave a Claude Code agent admin access to NIPRNet, an unclassified Pentagon network, to write and deploy code. It ran for two months, apparently sending login credentials, passwords and a network map to a third-party site NIST had flagged as high risk. Contractors must report breaches within 72 hours, yet sources say Accenture initially didn't report it at all. A source calls it "100 percent a national security threat."

ai_agentllm事件凭据泄露 @NetAskari source ↗

维护者基础设施成供应链单点,基础命令构建链的信任假设需重估

RT @S1r1u5_: What if you could backdoor Bash, ls, cp, rm, and other GNU software used by millions of systems? Our research led us to hack GNU Savannah, the infrastructure hosting their source code. The xz backdoor that never happened 🧵 https://t.co/0HPqlB1rZt

供应链后门基础设施 @0x64616e source ↗

补丁已进 AOSP 却未进公告,未打补丁的设备仍处窗口期。

RT @GrapheneOS: @IntCyberDigest Google committed a fix on June 15th and it was shipped in the security preview patches in mid-July. GrapheneOS shipped it in our security preview releases in July. It's in the upcoming November security bulletin but a patch was pushed to AOSP in October: https://t.co/h2cgMixDTG

移动端补丁android事件 @0xocdsec source ↗

路由层位于信任链中央,无需污染权重即可植入后门,自托管场景同样暴露。

RT @joshua_saxe: The security sensitivity of sites like OpenRouter is insane; imagine the damage you could do by inserting shell commands into people's token streams or just logging and ransoming people's data. Who needs to poison model weights when you can implant yourself within inference providers

ai_agentllm供应链 @0xocdsec source ↗

路由同时决定执行内容与审核逻辑,绕过点集中且难以审计。

RT @dyn___: @joshua_saxe Yep!! All of these model routers are instant RCE on basically millions? of systems. The models control what gets run, what gets displayed, the classifier checking the command... A good attack could be selective and only poison one command out of many or one system or everything.

ai_agentrcellm @0xocdsec source ↗

补丁仅在 macOS 27.2,未升级设备仍保留低门槛的本地提权路径

RT @Mas0nShi: Apple has patched a vulnerability I reported that had been present for over a decade. The fix was released in macOS 27.2. CVE-2026-112445 Of all the local privilege escalation (LPE) vulnerabilities I’ve seen, this is probably the easiest to exploit, the most straightforward, and the most bizarre—so bizarre that it makes you wonder whether it was an Apple backdoor. https://t.co/j5jcC1D6xd

lpecvemacos @ilove2pwn_ source ↗

攻击节奏整体抬升、应急响应资源排队,AI 被指加速了漏洞利用

RT @choblin29: WTF. Cybersecurity firm Sophos says its incident-response team handled SEVEN major incidents in SEVEN days, a workload that normally takes THREE MONTHS. The WSJ reports security firms are overwhelmed by what insiders call 'Black October.' MoxFive says newly breached companies may have to wait up to FIVE DAYS for help. Security executives say AI is helping attackers exploit vulnerabilities faster than defenders can respond.

事件报告 @0xocdsec source ↗

命令拼接注入的现成复现路径,凡调用 GPG 校验密钥的服务端都需自查。

RT @S1r1u5_: @rootxharsh @iamnoooob @HacktronAI Yep. Testing a GPG key put that email unquoted into a shell command. That gave us code execution as the web server user. With access to the database, we escalated our own account to superuser. https://t.co/f61MURAbzZ

rce提权命令注入 @0xocdsec source ↗
Recommended15

免费开源,覆盖防火墙、持久化与进程检测

RT @PabloSabbatella: 27 - Download the security tools from @objective_see see for MacOS. They are OpenSource and free, developed by @patrickwardle . - Lulu is a firewall that will protect you from incoming attacks and will stop malware from exfiltrating data if you get infected. - Knock Knock and BlockBlock protect you from persistent software - TaskExplorer analyzes all processes running, their signatures and checks with VirusTotal if they are infected. - ReiKey checks if some process is checking your keystrokes (keyloggers) - DoNotDisturb alerts if someone opened the lid from your Mac, takes a picture and sends it through Telegram - Netiquette shows you all incoming and outgoing connections and what process is doing each one. https://t.co/s1DkXM8ccP

工具移动端检测 @patrickwardle source ↗

研究者本人披露细节,修复版本明确为 Chrome 155,便于按版本排查。

RT @PaulosYibelo: Escaping Chrome's GPU sandbox through an undocumented macOS API. Discovered a sandboxed Chrome GPU process could inject keyboard events so sent ⌘+Space, typed "Calculator" & hit Return 😂. CVE-2026-106378 is fixed in Chrome 155. Update chrome dependancies. Happy Octoberfest.

浏览器逃逸cve @h4x0r_dz source ↗

绕过常见注入检测思路,PoC 已发布,可对照现有检测规则验证。

RT @MalDevAcademy: Released TpDirectCallbackHijack, a process injection PoC where you overwrite one function pointer inside an existing thread pool object (TP_IO / TP_ALPC / TP_JOB), post a single completion packet, and the victim's own worker thread dispatches your shellcode. Restore stub puts the pointer back on the way through, so the thread pool keeps working. No remote thread, no Get/SetThreadContext. Pool Party adjacent. Based on @_0xDeku's research. https://t.co/MzpJaCeFdh

工具poc绕过 @0x64616e source ↗

体系化的免费漏洞利用实验环境,适合验证与练手。

Want to learn exploit development for FREE? These VMs teach Linux privesc, stack/heap overflows, format strings, plus ASLR/NX/PIE bypasses. Nebula -> Phoenix -> Fusion. Built by Andrew Griffiths: https://t.co/B96HLdbbiF #ExploitDev #Pwn #CyberSecurity https://t.co/8OF326eai6

工具提权pwn @cr3ghost source ↗

按步可复现的取证路径,清历史并不等于无痕。

The employee cleared the browser history before handing the Windows laptop to the investigation team. At first glance, there was nothing useful left. No suspicious download entry. No obvious website in the browser. Then we checked the Windows DNS client cache. Windows temporarily stores recently resolved domain names so it does not have to query a DNS server repeatedly. In the image, the cache still contains https://t.co/aPxZiAauT3, its A record, the resolved IP address and the remaining time-to-live. This does not prove that the employee visited the domain in a browser. Another application, background service or security tool could have requested the same DNS resolution. The cache also does not identify exactly when the lookup occurred. But it gives an investigator a valuable lead that can be correlated with firewall logs, proxy records, process telemetry and DNS server logs. Browser history belongs to the browser. DNS evidence can exist elsewhere in the operating system. Useful PowerShell command: Get-DnsClientCache

取证windows事件 @Officialwhyte22 source ↗

本地部署模型的能力边界可能随之上移或收紧,直接影响自建安全 Agent 选型。

ok, i think the hammer is coming down very soon. my guess is that every chinese open model from now on will be severely nerfed on cybersecurity capabilities. i am not sure how well that is going to work, though. if you build an AI that can reason really well about code, can you actually remove its hacking capabilities without degrading its coding abilities? they might eventually realize that abliteration can restore those capabilities, and end up banning open weight releases altogether.

llmai_agent开源 @S1r1u5_ source ↗

覆盖从立足点到域控的完整杀伤链,可直接对照自查域内配置。

RT @_Karrab: Finally finished my Active Directory Pentesting Methodology! This can become your 𝗴𝗼-𝘁𝗼 𝗿𝗲𝘀𝗼𝘂𝗿𝗰𝗲 for AD pentesting: • 50+ Techniques • 300+ Commands • 100 Screenshots • Full kill chain (Foothold → PrivEsc → Lateral → Domain Compromise) https://t.co/sELxjC7HYf

工具ad报告 @0xocdsec source ↗

安全专用开源模型缺公开基准,加密思维链会限制二次分析与能力评估。

RT @bestswngs: 都在说 Deepseek 的cyber 模型,目前没看到bench的分数。 昨天做梦梦到有个鲸鱼娘和我描述了一下,我用 AI 重新生成了 HTML 然后截图,分数如下: 另外 Deepseek flahs cyber 这个模型还加密CoT了 😂 https://t.co/drbd0E8no4

llmai_agent报告 @0xocdsec source ↗
More86

补丁对比定位漏洞并给出 PoC,可作 Apple 平台研究范例

RT @Xplo8E: I've also written a blog explaining how I found CVE-2026-112621, from patch diffing to the PoC. This is my first write-up on Apple security research, so I'd really appreciate any feedback, corrections, or suggestions. It would help me learn and improve! :) https://t.co/NiDQeJlZoV

cvepoc补丁对比 @0xocdsec source ↗

形式化验证下沉到二进制执行层,长期影响漏洞可利用性判断

RT @v12sec: x86 formal verification is coming! The (first?) published machine-checked proof of a security property for the end-to-end execution of a Linux process running an x86 binary statically linked against glibc. And we believe it scales! A joint research project with @zellic_io 🧵👇 https://t.co/719dSrp0BG

形式化验证x86报告 @0xocdsec source ↗

可自托管的威胁情报平台,适合本地聚合 IOC 与 CTI 数据

opencti v7.261030.0 — Open Cyber Threat Intelligence Platform https://t.co/9rhCXm3CcK https://t.co/7l5u1EAcLZ

工具cti自托管 @KitPloit source ↗

12 days left until the 3-month deadline on my complaint to the Dutch DPA against HackerOne. They refused to answer a basic GDPR access request (Art. 15) after banning me. After 11 November I can take the matter further in court. Maximum possible fine for this type of violation: €20M or 4% of worldwide turnover. Still silence from both sides. I hope they continue to ignore this and I will fight in court for us.

gdpr披露事件 @k_firsov source ↗

Drunken Korean bar rant: Perhaps the most important takeaway from @unpromptedsf is the widening disconnect between demonstrated AI capability & the narrative being constructed around it. Frontier AI labs & their financial backers have a vested interest in weaponizing the credibility of exceptional researchers, turning legitimate breakthroughs into the perception that AI possesses capabilities it has yet to reliably demonstrate. Fear, hype, & the promise of technological superiority are powerful instruments for shaping public perception, influencing markets, & attracting investment. Yet much of this impressive research remains fundamentally dependent on years of human expertise, technical intuition, & highly specialized domain knowledge. These are not incidental contributions. They are often the very reason these breakthroughs are possible in the 1st place. The work coming out of labs like @calif_io is a perfect textbook example. Their research deserves genuine recognition & respect. Many of the people behind it are figures I have admired for years, as have countless others in our industry. That admiration, however, should not exempt the commercial interests funding & promoting their work from scrutiny. Their motivations, incentives, and ultimate objectives deserve far more attention. We should be able to celebrate meaningful advances in AI without allowing frontier labs to repackage exceptional human achievement as evidence of autonomous machine capability. Recognizing the value of AI should never require diminishing the expertise of the researchers making these advances possible. There is a fundamental difference between advancing science and manufacturing the perception of progress. Increasingly, the latter appears to be the more lucrative business. What concerns me most is how effectively this narrative has drawn some of our industry's most respected figures into its orbit, lending their hard-earned credibility to a marketing machine whose interests may not necessarily align with their own. And judging by the public reaction, it's working remarkably well... Sadly

llm议题 @Laughing_Mantis source ↗

AI 自动挖洞正在重塑漏洞发现与赏金生态

In >6 months cyber capabilities of OAI models are going to explode even more. Results? If you are a bughunter and hunt on OSS stuff (Chrome, Blockchain validators, Smart Contracts etc), you will get more and more dupes. The edge of "human element" will slowly vanish and models will finally reach a point where just pointing them at the codebase will spit out vulns like crazy (even the most obscure ones). Giga level scaffolding, domain specific harnesses, mega prompts, etc will lose their value , and will eventually become obsolete. The gap between private frontier vs Open weight models will widen further.

llmai_agent漏洞挖掘 @0xAsm0d3us source ↗

可否认加密实现示例,源码公开可供审计

went down a random rabbithole learning about deniable encryption systems and made a (probably not very secure because im not actually a crypto guy) thing (lots of clanking was involved) actual crypto people, pls tell me if (and why) its bad, thx https://t.co/28YsQcjGjO

工具加密 @codex_tf2 source ↗

iOS 27.0.1 上单个编码参数即可复现内核 panic,PoC 与日志已公开。

RT @MchinaMd57764: Reproduced CVE-2026-112705: AppleAVE2 kernel OOB write on iOS 27.0.1. Setting VideoToolbox NumberOfTemporalLayers=16463 panics the kernel on first encode. Fixed in iOS 27.2. PoC + panic log: https://t.co/NmmoOyVBOi Found by Calif / Chollima team — independent repro. #iOS #CVE https://t.co/5DAKSwgtWB

cve内核poc移动端 @0xocdsec source ↗

提示 CVSS 与实际可利用性脱节,并给出具体反例。

Many of the CVEs you see coming out these days are actually not very weaponizable against real production systems, for one reason or another Example: Anthropic’s recently reported Linux network RCE, which takes almost an hour to run, causes a kernel panic on any system that has even the default level Linux Firewall, in our testing Often, the reason why you have even of these these bugs is because firms put them in the throwaway conference-talk category Important to keep this in mind

rce内核报告 @C2IRIS source ↗

点名窃取 ChatGPT/Claude 会话令牌即等同全权限,与 Agent 场景直接相关。

initial access will do most of the work these days. 1. initial access in application level bug would dump crown jewels, most of the monitoring wouldn't catch application level bug exploitation atleast in my experience 2. initial access in victim computer would let access to almost anything, i am very suspicious of any current endpoint security tools working well 3. compromise victim browser, game over, get https://t.co/MNUWpADcdL or https://t.co/CZ5gT88CXx tokens you have everything the list goes on. i have never saw a lateral movement making our life harder, it is always the initial access that is pain

ai_agent提权事件 @S1r1u5_ source ↗

给出具体持久化路径 com.apple.sysmond.plist,可直接写成检测规则。

RT @objective_see: vs. our free, open-source security tools: 1️⃣ BlockBlock alerts when the payload persists via ~/Library/LaunchAgents/com.apple.sysmond.plist 2️⃣ LuLu alerts when the payload attempts to connect to its C&C server via SSH and/or HTTP Tools! 🛡️😇 https://t.co/7zBX8pcMrO https://t.co/vxPzOmBdvW

工具移动端事件 @patrickwardle source ↗

说明雇佣间谍软件攻击仍在活跃投放,移动端高价值目标威胁未减弱。

RT @craiu: Heads-up: New round of Apple Threat Notifications is going out right now, distinct from the August wave and tied to new attacks! Apple’s notice: “Today’s notification is being sent to targeted users in 99 countries, and to date we have notified users in over 150 countries in total. The extreme cost, sophistication, and worldwide nature of mercenary spyware attacks make them some of the most advanced digital threats in existence today. As a result, Apple does not attribute the attacks or the notice you’re receiving to any specific attackers or geographical regions.” Received one and need help? DM me, happy to assist.

移动端事件间谍软件 @0xocdsec source ↗

面向多架构二进制分析的新开源工具,作者此前有 Android 反编译工具积累。

RT @Farenain: As promised, and thanks to @yates82 and @jeppojeps for the help, here you can find Dragon-tales: https://t.co/nKayn15Fr4 I hope the tool can be useful for someone else, and we will continue improving it with support for different architectures. https://t.co/AE4mqDhEaJ

工具逆向 @alkalinesec source ↗

给出防守资源投放的优先级判断,与常见的入口加固叙事形成对照。

Most of the cybersecurity world is focused on the risks and vulnerabilities that lead to initial access, but defenders get *way* more leverage from making lateral movement impossible. You want to get to where the most an attacker can do is screen-scrape the victim's workstation.

横向移动防御报告 @dinodaizovi source ↗

若成立,意味着最新 iOS 版本存在可被利用的完整链,但尚无细节与复现材料。

RT @ikotas00: We got a jailbreak working on iOS 27.2. Eating the Apple first feels a lot like The Very Hungry Caterpillar. https://t.co/IOA0BOYc0T

移动端越狱poc @0xocdsec source ↗

硬件钱包供应链植入后门,需换新助记词转移资产方能止损

RT @MagicalTux: Ledger isn't the only one, @Trezor devices sold by CryptoBilis also had an implant. IMPORTANT: If you bought a Trezor from CryptoBilis from March 2026 onwards and are using it to store funds, please move your funds to a newly created wallet with a new seed as soon as possible. If you have a device you are unsure of, DM me. I'll happily cover shipping fees and analyze your device. https://t.co/M88lXN0E6x

供应链固件后门 @0xocdsec source ↗

第三方脚本是常见盲区,与自动化 Agent 扫描叠加后攻击面更大

RT @aaronjmars: 2.7M drained because they've used a third party script (debugbear) to analyze their website speed the cyber apocalypse is not just about CVEs & vulnerabilities, the main risk is chaining vulnerabilities across supply chain attackers have unlimited time, as time scale with more agents the moat becomes the agent framework & the surface of what you can scan

供应链事件web @cyb3rops source ↗

I definitely agree with @chrisrohlf here, but this logic shouldn't stop at libraries. if the argument is that dependencies introduce too much risk, then you should also be looking at rewriting gnu coreutils, bash, and other foundational infrastructure. that is not happening anytime soon, though. It seems to me we need to throw an absurd amount of tokens at securing this infrastructure and helping its maintainers, you still probably not solving insider threat that could lead to things lik xz backdoor. and if you want to look into them, please, don't just dump vulnerabilities on already overwhelmed maintainers, provide patches always i also think @AnthropicAI and @OpenAI should invest heavily in securing this attack surface. i have reached out to some and proposed focusing more on this infrastructure, but unfortunately, i haven't seen much interest.

供应链讨论 @S1r1u5_ source ↗

主流 AD 攻击工具默认转向签名与通道绑定,攻防口径同步变化。

RT @_dirkjan: Last week me and GPT-6 wrote a wrapper library for ldap3 AD auth methods and ported several of my AD tools to use it. BloodHound-py, krbrelayx and adidnsdump now use it for their LDAP connections: signing/sealing over 389 by default (ntlm+kerberos) and channel binding for ldaps.

adldap工具绕过 @_EthicalChaos_ source ↗

模型代际差异直接影响攻防效率与检测策略选型。

RT @S1r1u5_: >But there’s also an attitude that new model capabilities are “no big deal” that sometimes gets packaged with that, and I very much don’t agree with that. yea, i don't agree with that too. using these models every day for cyber, the difference between models versions are staggering. and i strongly think it is a big deal because if you are a defender you need to make sure you have the current frontier.

llmai_agent威胁情报 @S1r1u5_ source ↗

@ProtonPrivacy Being a long time supporter of Proton and its services I am really disappointed to hear that your CEO is having such a laissez-faire attitude to product security when it comes to AI involvement as stated in the @404mediaco report. I am not demanding that the company is not using AI in their products at all but it feels this 'move fast and break things' attitude is just not going to work well for a privacy and security-first marketing line that Proton has been successfully selling in the past.

llm隐私观点 @NetAskari source ↗

说明权限禁用并非有效边界,同类传感器侧信道可复用于其他 SDK。

开发者使用 AI 分析 #百度网盘 iOS 版为什么禁用运动传感器权限还能跳转 #摇一摇 广告:因为百度网盘还开发了磁力计框架。 百度网盘 iOS 版广告框架里有个名为 NoShake (没有运动) 的变体,该变体会检测苹果 Core Location 框架中的 CLHeading 获取设备方向数据。 该数据可以提供设备相对于磁北方向的朝向和用于计算方向的三轴磁场数据。 简单来说,手机姿态发生变化时磁力计数据也会随之变化,这样百度网盘还是能判定用户摇晃了手机进而跳转广告。 AI 分析报告还提到了零点击自动跳转广告,当特定参数组合满足条件时,等待约 1 秒后即便用户没有点击,SDK 也会自动执行广告点击流程。

移动端绕过逆向 @landiantech source ↗

迄今最大规模加密盗窃,跨 10 条链二十分钟内完成转移,资金漂移手法可参照。

RT @AZ_Intel_: NEW: North Korean state-sponsored hackers stole approximately $6.36 billion from Binance in the largest cryptocurrency theft on record, according to blockchain intelligence firm TRM Labs, which attributed the attack to the TraderTraitor group. The funds moved across 10 blockchains beginning at 20:05 UTC on October 29, with most transfers occurring within 20 minutes. The attack bears similarities to the February 2025 Bybit hack, also attributed to TraderTraitor by the FBI. Tether, Circle, Ripple, World Liberty Financial and Arbitrum’s Security Council have frozen approximately $2.7 billion, while an estimated $2.4 billion remains beyond their reach.

事件加密货币apt @0xocdsec source ↗

系列博文附可编译示例,适合理解样本加密与对抗分析的具体做法。

https://t.co/7Og7jNtx5a next one from my blog #cryptography series. enjoy! https://t.co/A0sXpk968f #hacking #malware #cybersec #cybersecurity #threatintel #research #ML #AI #programming #math https://t.co/9byAbiWCqr

恶意软件密码学报告 @cocomelonckz source ↗

刻意的可持续收割设计,说明攻击者意在长期驻留而非一次性清空。

RT @uttam_singhk: The attacker left Binance wallet exactly 1.00 ETH after every sweep. so that leaves enough gas for binance's normal withdrawals to keep running. The wallet stayed operational and kept receiving deposits the attacker could take again 🥲 https://t.co/8xuTjacJMm

事件加密货币 @0xocdsec source ↗

增量迁移比全量重写更可落地,可作为内存安全改造的路径参考。

RT @calif_io: Just in time for Halloween: Rust in Peace 🎃 We often share how we break software. Today, we show how we help defend it, replacing vulnerable C++ code with Rust, without a full rewrite. https://t.co/Gs2ExequYl https://t.co/K0jkk8n3Gz

工具内存安全 @Teach2Breach source ↗

供应链脚本与人员入口仍是实际突破口,可作为 agent 供应链风险的对照

i would have said "told you so!" but i would be surprised if any of this was due to AI. this is just typical crypto security. polygon got hacked because of its reliance on some insane third-party js? di you guys learn nothing from the slope wallet hack? and binance, employee got compromised? we have seen this in countless crypto hacks before. i am not sure about the initial compromise, but i would be surprised if ai played a role rather than some good old-fashioned, carefully crafted supply chain attack or phishing. "make one mistake and lazarus will be up your ass" was true long before ai.

供应链事件 @S1r1u5_ source ↗

模型训练数据来源转向专业代码,直接影响代码审计与漏洞发现能力边界

OpenAI is reportedly offering huge money to buy out smaller HFT firms and companies like [redacted], because they realize that they really do need more data that is handcrafted, high quality, high performance code in order to improve coding performance of the models. GitHub slop, vibeslop synthetic self-licking-ice cream-cone data and even what many AI-bros wrongly thought of as high quality open source code (like the Linux kernel ~lol~) is not improving the quality of the code that their models are producing. The reward-hacking behavior towards “just make the code compile to earn your reward” has been proven to be extremely counter-productive. Current SOTA models struggle to maintain coherent awareness of critical performance concepts like cache locality, branch prediction, instruction-level parallelism, memory ordering, false sharing, lock contention, vectorization, NUMA topology and the actual cost of moving data through a modern memory hierarchy. People will tell you that this stuff doesn’t matter, because modern compilers are so good. And in a normal world, I would concur with this sentiment. Normally, it would be true. In today’s world, you don’t need to really panic about a few too many if statements, because compilers will optimize away the cost. But these models, some how, some way, have an uncanny ability to write code that is SO bad, that not even the compilers can fix it. Their understanding of I/O performance is arguably even worse. They routinely generate code that is almost completely oblivious to storage hardware: pathological write amplification, excessive fsync() calls, unnecessary serialization, unbounded logging, tiny synchronous writes, redundant disk operations and utterly incompetent batching and buffering strategies. Under sustained high-performance workloads, these pathologies can go well beyond merely degrading throughput. They can generate enough unnecessary physical I/O to accelerate SSD wear, exhaust write endurance, and contribute to premature drive failure. I’ve recently experienced hardware failure myself when I stupidly trusted GPT-6 to write and run a high throughput data service. Which leads one to wonder… with trillions of tokens’ worth of newly vibecoded, I/O-oblivious software now finding its way into production environments, including the major hyperscale clouds, has anyone actually examined whether this phenomenon is beginning to materially affect storage-device failure rates? And this is to say nothing of their performance at higher levels of abstraction. The conventional wisdom is that these models are much better at Python, JavaScript and the various higher-level frameworks than they are at systems programming. People who spend most of their time in that world will even confidently proclaim that “coding is solved.” But consider the extraordinary irony that OpenAI’s own ChatGPT/Codex desktop applications, built around JavaScript frameworks, have been the subject of repeated, frequent, catastrophic resource-management failures. Most commonly, runaway memory consumption reaching tens of gigabytes, psychotic background I/O loops, and SQLite logging behavior that has generated tens or even hundreds of gigabytes of disk writes during ordinary application use. And again - this is from the application written with the benefit of OpenAI’s own internal models.

llm训练数据 @C2IRIS source ↗

RT @EpochAIResearch: The computing power of AI chips sold each quarter has grown 2.5× per year since the start of 2024. That’s down from roughly 6× in 2023, led by Nvidia’s then-new H100 GPU. Some evidence suggests growth slowed further. But quarterly sales are noisy, and our data ends just as Nvidia begins shipping its next-generation chips, a move that has historically spiked sales. This could be a lull between generations. See our website for more details on the analysis: https://t.co/rcxMLsJpcS This is from our AI chip sales dataset, which uses financial reports, company disclosures, and more to estimate compute, power usage, and spending for a wide variety of chips: https://t.co/t5f7LRClAI

ai芯片报告 @0xocdsec source ↗

Has been for at least the last 20 years, but anti-cheats are the most sophisticated they have ever been now, and methods of cheating are too — so, what’s your point? I’m not suggesting don’t fight it, to be clear, but I don’t think anyone is coping about rampant cheating. I also think your # is a little low… it’s closer to $​10bn. I think the biggest issue is these studios that lack the staff keep trying to roll their own anti-cheat so they can keep everything in house; and they’re the easiest targets for cheat providers to make a hefty bag and bug out before they contract a useful middleware AC. There’s a new, popular game that is suffering from this issue right now.

反作弊游戏 @daaximus source ↗