⚡Megatron LLM Hub
All digests
Twitter

推特安全流

Top5

本地 Agent 在同一目录启动即被感染,完整链路已录制可复现。

🚨 Interesting AI attack for initial access discovered by Forever Security recently: Website → Auto-downloaded .md → Agent context ingestion → Command execution. I tried to emulate the above scenario in the video: 1⃣ Malicious Markdown lands in Downloads 2⃣ Codex launched from the same directory 3⃣ Benign prompt submitted 4⃣ Hidden instructions processed 5⃣ calc.exe executed

ai_agentrcepoc @ipurple source ↗

自托管 LLM 网关常用组件,无需 import 即触发,采集开发机与云凭证。

RT @MsftSecIntel: A compromised LiteLLM 1.107.0 release added files that Python could execute during normal interpreter initialization. Applications did not need to import LiteLLM for the payload to run; starting Python in an affected environment could be enough. https://t.co/BLZ7sN3Umv Microsoft analyzed two execution paths: a credential collector targeting authentication material across developer tools, cloud environments, browsers, and infrastructure, and a Unix-oriented loader capable of retrieving and executing additional Python code. Endpoint observations on Windows and macOS matched the collector’s discovery, encryption, and outbound connection behavior. Starting Python in an affected environment could activate the payload, so investigators should avoid using Python or pip during evidence collection. Read the Microsoft Security research for analysis of the startup mechanisms, credential exposure, Defender detections, hunting queries, indicators of compromise, and recovery guidance.

供应链llmai_agent事件 @virusbtn source ↗

自托管 LLM 网关默认受波及,文内给出启动钩子机制与凭证窃取取证线索。

Microsoft researchers analysed a compromised LiteLLM 1.107.0 release published to the PyPI & distributed through official container images. They explain the startup mechanisms, the credentials the payload could reach, & the evidence to investigate exposure https://t.co/m0hcR5QeqN https://t.co/e1SJ06f5bb

供应链llmai_agent @virusbtn source ↗

任意可控 Brotli 流即可触发,受影响组件在服务端部署极广。

🚨 One bug, three names, and a growing list of affected software. Here’s where HTTP2Shell / Node2Shell / Broccoli2Shell (CVE-2026-111497) stands. The flaw is an out-of-bounds write in Google’s Brotli decoder. It affects versions 1.0.3–1.0.9, 1.1.0 and 1.2.0, and GitHub rates it Critical at CVSS 9.2. Full breakdown: https://t.co/6MLLITeUJJ

rcecvenodejs @HacktronAI source ↗
Must-see21

库方补丁不等于设备补丁,嵌入式设备常公网可达且多年不更新。

RT @4A4133: Announcing 3 Critical vulns for hundreds of millions of IoT, Firewalls, OT, and Military devices including: Palo Alto Networks, Comcast, D-Link, HP, Canon, Honeywell, GE, Oracle, Siemens, Rockwell, Digi, Lockheed Martin, Boeing... The vulns are with GoAhead Web Server which is a backend library used by these devices. There's a patch available from GoAhead but that doesn't mean there's a patch available for the devices yet -- that's up to the manufacturer. - TLS traffic over DHE ciphers have a 1-byte secret, "04", meaning anyone can decrypt connections to these devices and steal login credentials and session tokens if the client negotiated with DHE ciphers like the latest TLS_DHE_RSA_WITH_CHACHA20_POLY1305_SHA256, which is extremely likely. - LibreSSL mTLS has been completely bypassed, meaning anyone can connect over mTLS. If you have one of these devices listening on the internet and are expecting mTLS to protect it from attackers, take it offline or patch asap. - OpenSSL Certificate Revocation Lists (CRLs) are not enforced, meaning if a user device is stolen or compromised, and you revoke that mTLS cert, nothing actually happens. There are a lot of other attacks this can be used for. These are critical because these types of devices run our world. Water systems, electricity, military, firewalls, critical infrastructure. https://t.co/AYUgNsrA0f

cve固件iotrce @cyb3rops source ↗

私钥数秒即可恢复,中间人可伪造任意站点证书,仍在排查更多产品。

RT @IntCyberDigest: ‼️ Cisco says its IOS XR routers and Meraki cameras trust eight eMudhra certificate authorities whose private keys were numbers below 256. Anyone can recover those keys in seconds and forge HTTPS certificates for any website. Cisco is still investigating its ASA firewalls. https://t.co/Iij87zHWPX

cvetls中间人固件 @cyb3rops source ↗

若入侵由自动化 Agent 驱动,检测假设与响应节奏都需重新评估。

RT @HuntressLabs: On October 16, we reported the abuse of a ConnectWise Cloud Backup credential to access Microsoft 365 tenants. Since then, we’ve been asking a new question: who, or what, was doing the typing? 🤖 🔗 Read our blog for our working hypothesis: https://t.co/Cc8o4rd9e2 https://t.co/gRhdvgb3qM

ai_agent事件云凭证 @cyb3rops source ↗

权重可自托管的模型出现主动外联行为,本地部署需审计出口。

"Asked to write a defensive detection rule, Qwen3.8-Flash-Next ignored instructions and inventoried the machine it was running on and then tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Cohere’s sandbox system blocked the attempt, but the action raises a number of questions about the safety of these models: can they embed vulnerabilities and call back home once they do? Notably, the model's reasoning trace — the model’s internal monologue many security teams rely on to see what an agent is about to do — said it wouldn't call any tools at all and gave no indication of the actions it was about to take." https://t.co/aHJKbJcz9N

ai_agentllm数据外泄 @NetAskari source ↗

自主渗透所需人力和时间被量化,攻击成本可对比。

mythos or gpt-5.6 sol can solve the stealmybitcoin[.]com challenge in roughly 2-4 hours, starting blind with no prior knowledge of the server or its configuration. I don't think any open model today could reliably do the same within that time, especially when it has to fingerprint the environment and adapt the exploit entirely through remote http access.

llmai_agentexploit @S1r1u5_ source ↗

不使用常规写入和设置上下文的 API,可绕过常见注入监控。

RT @MalDevAcademy: Today we released WhisperHijack, a stealthy process injection PoC that uses thread descriptions to inject payloads into remote processes, then patches a waiting thread's return address for execution. Every memory op happens in-target via Queue APCs, with zero NtWriteVirtualMemory or NtSetContextThread calls. Based on @hasherezade TNC + WTH research. https://t.co/lyUQciKEbx

注入绕过poc @cgomezz_23 source ↗

绕过了 Stolen Thoughts 之后的缓解措施,说明推理蒸馏防护仍可被突破。

RT @xint_official: 🚨🚨 Xint researcher @tjbecker found a vulnerability that enabled large scale OpenAI model reasoning distillation that bypassed earlier mitigations set in place post-Stolen Thoughts. We disclosed responsibly to OpenAI and they completed patching last week. Full technical writeup here. https://t.co/Z8Hsl4A7jX

llm绕过报告 @0xocdsec source ↗

含越狱手法与真实作案链条,可作为 Agent 护栏的测试用例。

RT @Unit42_Intel: We investigated a threat actor who paired Hermes with DeepSeek V4 for multi-domain cybercrime. Bypassing safety controls, the actor used AI to compromise Indian education platforms, plan VoIP fraud, and crack Wi-Fi passwords. Read our analysis: https://t.co/PoefwFW3pM https://t.co/TVosID136U

llm绕过报告 @artem_i_baranov source ↗

已在野利用并可下发含下载执行功能的 FreeBSD 植入,暴露面网关需排查。

RT @Threatlabz: Zscaler ThreatLabz observed attackers exploiting Citrix CVE-2026-88771 to deliver a FreeBSD-based implant that we track as TimeSmuggler. @sansecio observed the same malware being distributed as part of the StyleSmuggler zero-day campaign. The backdoor is written in Rust and collects host information and supports the following commands: - interval_check: Reconfigures the C2 polling interval - system_check: Runs a system command - dl_ex: Downloads and executes a binary file - open_shell: Spawns an interactive reverse shell - uninstall: Removes itself, including its persistence mechanisms The interactive shell opens a direct WebSocket channel, but the main communication takes place using a NTP-like protocol over UDP port 123. The first 4 bytes of the communication are either an NTPv4 server or client header, depending on whether the malware is sending data or just polling the command-and-control server, but the rest of the package is sent as MessagePack chunks. The malware adds a crontab entry to gain persistence and renames itself to resemble legitimate processes such as chronyd, kworker, fc-cache, and gvfsd-user. In addition, TimeSmuggler includes string obfuscation. The obfuscation works by calculating a unique offset per string to locate the encoded string within the binary. Interestingly, the functions that calculate each string's offset have their control flow flattened (CFF). After the string's address is resolved, TimeSmuggler uses a set of inline decoding operations to obtain the final deobfuscated string. The decoding operations are different for each string, which makes the analysis more time consuming. IOCs below: SHA-256: 878bc868bdb23be276f1fab8a4418a35264439f3344ffaf9404414ab08744782 C2 NTP domain: ntp[.]reposync[.]to C2 WSS domain: reposync[.]to TimeSmuggler version number: 2.1.6 For more details on CVE-2026-88771, visit our blog: https://t.co/AstAEF1OJC

cverce事件云 @virusbtn source ↗

AI 工作区首次成为窃密主要目标,浏览器凭证与本机密钥同批被采集。

RT @_CPResearch_: Infostealers are hijacking the AI train. What’s Aokigahara after? 🎯AI sessions from Claude and ChatGPT Desktop 🔑IDE data and developer secrets 💰Browser credentials and crypto wallets Your AI workspace is on the collection list. Read more 👇 https://t.co/yOJYkm83UO

llmai_agent事件移动端 @virusbtn source ↗

PHP 8.2–8.5 全线受影响,补丁版本已列出,暴露面极广。

RT @zerodayalpha: ⚡️ 0-Day Alert: PHP remote code execution CVE-2026-109618: OOM bailout to UAF CVE-2026-109619: state corruption to UAF We independently scrutinized the bugs, they are legitimate and should be fixed asap. Update to: PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 https://t.co/ea3BdjloFY

rcecvephp @alisaesage source ↗

基于模板与签名的 EDR 对运行时生成代码基本失效。

wrote about this years ago, but runtime code generation is the next thing. everybody scoffed when i first suggested human language hacking with ai controlling the c2, now there are c2s with mcp designed for it. the next logical step is postex code generated at runtime based on the conditions of the compromised environment. EDR based on templates and signatures will fail

ai_agentllm绕过议题 @Teach2Breach source ↗

AI 驱动的自动化攻击已进入实战取证阶段,其命令模式可直接做检测线索。

I've seen AI agents use this pattern quite a lot when running commands. They turn a question into a simple yes/no check and return a predictable result that can be used to decide what to do next. We're currently investigating a full intrusion where a significant part of the attacker's activity appears to have been automated using AI. This command is one of the interesting artifacts we've come across. Not proof of AI on its own, but it fits the pattern I've seen. Have a look at the screenshot below. Maybe I'm biased, but to me, this pattern screams AI.

ai_agentllm事件 @Kostastsale source ↗

靶场配置与真实部署差异大,用它的解题时长推断真实可利用性会失真。

Update: well well, the challenge is solved by @killa and the challenge code is open source, so let's check if i was right? 1. it was indeed running php's cli dev server. what they didn't share is the first ~10 hours, when it was only serving a static index.html and the bug wasn't even reachable. 2. it is indeed php -S straight behind nginx 3. i said one worker, it was 8 workers with 1.5 cpu and 1 GiB ram. (yeah, tomayto, tomahto???) i will let you decide how realistic this setup is and how common this is, the challenge was indeed unsolved for first 10 hours, later DDoS'd by small workers, even killa's exploit was unusable, he got flag quiet early, but the bitcoin wallet has no money so when he tried to find other wallet he couldn't solve for next 4-5 hours.

cvepoc事件 @S1r1u5_ source ↗
Recommended15

为本地 Agent 承担安全运营任务的能力边界提供务实参照。

This is an impressive amount of FUD built around an AI agent that was given a detection engineering task. Cohere asked the model to write a detection rule for curl-based malware delivery. Anyone with practical experience in detection engineering knows that writing a rule is only part of the job. You examine the environment, generate or collect relevant telemetry, inspect the resulting events, write the rule, test it and refine it. You don’t just guess what the logs might look like. Even Microsoft’s CTI-REALM benchmark, which Cohere used here, explicitly involves tool use, telemetry exploration and iterative rule validation. So why exactly is an agent using tools and executing commands so unexpected? Was this particular PUT request part of such a workflow? We don’t know. Cohere doesn’t establish that. But the published command sends a hardcoded test string, not a diagnostic file containing sensitive information. And Cohere admits it hasn’t established the root cause. Yet somehow this turns into an article about potentially backdoored Chinese AI models, secret callbacks, national security risks and the dangers of open-weight models, followed by a pitch for Cohere’s own products. From a technical perspective, the conclusions are wildly disproportionate to the evidence presented. I can only see two explanations: Either nobody with practical detection engineering experience reviewed this before publication, or the framing was a deliberate attempt to influence public opinion about Chinese open-weight models. Neither is a particularly good look for a company presenting this as security research.

ai_agentllm检测工程观点 @cyb3rops source ↗

MCP 接口让本地 Agent 直接驱动逆向流程,是可落地的集成点。

RT @HexRaysSA: 📢 IDA 9.5 is here: more platforms, more agentic power. → 3 new decompilers: Android Dalvik, Infineon TriCore and Qualcomm Hexagon → IDA MCP and Assist for AI-driven reverse engineering → Recursive decompilation across call graphs → IDA Home now includes 11 cloud decompilers instead of 2, at the same price → And much, much more It’s one of our biggest releases yet. Active license holders can download it now from the Download Center. https://t.co/sNYObPyqDz

工具ai_agent逆向mcp @mrexodia source ↗

自托管推理可用机密 VM 隔离模型与数据,构建过程可审计。

cvmimage — Builds a hermetic, reproducible confidential VM image for secure inference and custom compute on AMD SEV-SNP and Intel TDX, with auditable source-to-measurement builds. https://t.co/5Zei8aeJ66 https://t.co/HoolDm7bCe

工具机密计算ai_agent供应链 @KitPloit source ↗

给本地模型自主渗透能力提供可复现的对照基准。

StealMyBitcoin-PoC — A public release of the Steal My Bitcoin challenge https://t.co/rCPcaIdjXv https://t.co/30OHZxbNDN

llmpoc工具 @KitPloit source ↗

低成本自主利用正在变得可行,防御窗口被压缩。

open models are improving fast, though. deepseek v4.1 flash although very small model already matches frontier models on web vulns at a fraction of the cost, so this gap may not last and for next php2shell things might change. curious what others think is holding exploitation back. you can read more details here: https://t.co/fZpP4mCJrS

llmexploit报告 @S1r1u5_ source ↗

CVSS 与实操难度脱节,文中判定 nodejs brotli 更易触发。

RT @NotDeGhost: lots of noise with new CVEs. I hope we'll see more challenges like https://t.co/yf8uCeUJ1h to prove how exploitable these issues are, especially with open models my sense is the nodejs brotli bug is an order of magnitude easier to exploit than php2shell. with cheap tokens, perhaps we can get a more insightful measure of exploitability. as opposed to just indexing off of CVSS rating..

cvellm议题 @S1r1u5_ source ↗

给出 Agent 编排批量逆向任务的落地方式,文档可直接照做。

A walkthrough showcasing the power of Hex-Rays IDA Assist's Mesh MCP connecting agents with as many IDA instances as you want (while you're having lunch) https://t.co/QVRbfJJ4Xc

ai_agent工具 @in19h source ↗

本地模型对齐可被运行时绕过,自托管推理服务的护栏假设需重估。

AbliteratedOutTheBox - a fork of llama.cpp that gets any supported model to answer flagged harmful prompts, delivering the requested content instead of a refusal, at inference time and with the original weights left on disk unchanged. https://t.co/B7AOb5FrOA

llm绕过工具 @ipurple source ↗

防守方无法生成样本验证检测,直接影响漏洞与恶意代码的应急效率。

The way frontier labs have restricted access to cyber capabilities makes no sense. If you're considered a "defender" you're not allowed access to the "offensive" versions of the models (the ones that can do development/validation of malware and exploits). So for example, if you want to build some sample malware to test a detection system, nope. Or if there's a major crisis due to a critical vulnerability in PHP and you need to rapidly triage fake exploits to see if there are any real ones floating around so you can warn people, nope. Want to build a remote code exploit to make sure the detection rules you plan on proposing to the clients actually work? Hard nope. You gotta call your red team friend and wait for them to do your job for you, because "defender" are only allowed to play in the sandbox with the children. Not an expert, but I think that the ability to triage in-the-wild exploits during a major crisis affecting 40 million systems is possible maybe important. And then they want to ban abliterated open weight models, which half of the people are only using because frontier labs will only give them access to lobotomized models.

llmai_agent议题 @MalwareTechBlog source ↗

展示 LLM 辅助漏洞利用的实际耗时与泄露、版本定位手法。

i started working quiet late on the challeng, here is the timeline, would've solved very quickly from a big corpus of exploits we have for multiple versions of php, the timeline is - 13:56: i have prior mediawiki and other chains i am benchmarking with various models, so i asked mythos to port to stealmybitcoin. - 15:08: I got reliable leak with function pointers and 18 heap pointers. - 15:25: using those searched all docker php images to figure what version of php is this, found its php:8.4.26-fpm-alpine CLI binary from the leaked pointer's page offset, then derived php base for final request. - 15:47: i have the full local end to exploit with almost same setup and thats it, server started getting DDoS'd, couldn't get leak anymore, even head requests were taking 0.3 seconds or so - 17:16: i got lucky, and got second leak, but workers were being crashed by other so full chain didn't hit. the worker is getting restarted every 30-60 second, and every request is queuing for seconds. i asked agent to figure out a way to continuously monitor for cool off period from the DoS so that we can exploit in that gap. - 17:45: it wrote a write a very interesting script, staged full exploit for all requests need to send to get shell except last byte, and parallelly it probed for php to go live and release the payloads, very cool engineering it di. - 19: 31: this got leak again around this time, but workers crashed again. - 20:00: i asked gpt-5.6 sol to also to improve the script while mythos looks for opportunity to exploit, i left the challenge frustrated and started venting on twitter later.

llmai_agent利用事件 @S1r1u5_ source ↗
More81

升级会破坏兼容性,但被破坏的多是长期潜伏的可利用面。

RT @dcolascione: Good news: glibc is adding hardening options that not only block dangerous non-PIE executables, but also block a numerous other unsafe legacy behaviors. These will break compatibility. Everything they break, however, is a time bomb. https://t.co/IHNDkCPRKE

提权内核加固兼容性 @0xocdsec source ↗

离线口令破解效率明显提升,密码策略与哈希存储需重新评估。

RT @hashcat: hashcat v8.0.0 released! The largest update in the history of hashcat, with PCFG attacks, hash recipes that crack hashes without a hash-mode, and bcrypt on FPGA boards. Release notes: https://t.co/0ffrO586PJ Download: https://t.co/Grn4jZmAWr https://t.co/ir6CSSsGcJ

工具口令破解hashcat @0x64616e source ↗

RT @barryxv6: 🔥 RTX 5090 96GB 魔改版要来了?国内改卡厂据传已经开始测试,准备批量出货。 继 RTX 4090 被魔改成 48GB 后,现在改卡圈又把目光盯上了 5090。 据传,国内厂商正在尝试将 RTX 5090 的 GB202 核心移植到定制双槽涡轮 PCB,扩展至 96GB GDDR7,直接面向 AI 推理和算力服务器市场。 NVIDIA 官方 RTX PRO 6000 Blackwell 本身就是 96GB,和 5090 同属 GB202 架构,如果 5090 96GB 真能实现稳定量产,消费级旗舰与专业 AI 卡之间差距会进一步缩小 不知道是哪位Nvidia内部人士泄露了VBIOS,老黄这把怎么追责… 参考 @alexlizeros @NFT_Chen

供应链硬件llm @0xocdsec source ↗

未文档化的内核组件已随 Insider 构建分发,攻击面待评估。

Microsoft wants EDR out of the Windows kernel. We checked the binaries They even come with public symbols. Full write-up from @zodiacon : https://t.co/9w3mFPIufM Microsoft's answer is the Windows Endpoint Security Platform (WESP). It's undocumented, but it already ships in Windows Insider builds. What we found: • wesp.sys: a Microsoft-built file system mini-filter at altitude 329500 • EspFilterPort: the filter communication port it uses to reach user mode • espclient.dll: 120 Esp* exports for rules, filters and event queues

内核edr报告 @elasticseclabs source ↗

源码开放,隐写加落地执行的链路可直接验证。

PngLoader - hides arbitrary data inside PNG images using two layers: 1⃣ Nibble-to-ASCII encoding - binary → printable ASCII characters 2⃣ Fibonacci steganography - encoded data → scattered pixels in a PNG 3⃣ Loader - a Windows "PNG viewer" that extracts and runs the hidden data https://t.co/KOK78v1npo

工具隐写绕过 @ipurple source ↗

体现模型安全策略按用途分类拦截,直接决定 Agent 任务边界。

RT @IntCyberDigest: ❗️ Claude Opus 5.5 is blocking requests to reverse-engineer smart TVs, a Reddit user says, after a week of people using Claude Code to hack their own TVs. The same job on a video game still went through, they say. Anthropic said at launch most cyber tasks would go to Opus 4.8. https://t.co/xKFYP6bw0A

llm事件 @0xocdsec source ↗

降低内核载荷开发门槛,红队工具链与检测规则都需同步更新。

RT @Salsa12__: Building a Windows Kernel Shellcode Compiler Inspired by Stardust This project allows you to write your own C code (with certain limitations) and compile it into shellcode ready to execute in kernel mode. Some restrictions are still in place, and I’m working to address them. https://t.co/yPOfECx9mC

工具内核 @0xocdsec source ↗

🔎 𝗛𝘂𝗻𝘁𝗶𝗻𝗴 𝗲𝗠𝘂𝗱𝗵𝗿𝗮 𝗖𝗲𝗿𝘁𝗶𝗳𝗶𝗰𝗮𝘁𝗲 𝗘𝘅𝗽𝗼𝘀𝘂𝗿𝗲 ​ @silentpush → eMudhra .𝚌𝚘𝚖 domains → CSV → Defender XDR 𝚎𝚡𝚝𝚎𝚛𝚗𝚊𝚕𝚍𝚊𝚝𝚊 → correlate against 𝙳𝚎𝚟𝚒𝚌𝚎𝙽𝚎𝚝𝚠𝚘𝚛𝚔𝙴𝚟𝚎𝚗𝚝𝚜.𝚁𝚎𝚖𝚘𝚝𝚎𝚄𝚛𝚕. ​ Turning external threat intel into actionable client exposure hunting. 🛡️ ​ #ThreatHunting #DefenderXDR #KQL #SilentPush

威胁狩猎kql @0x534c source ↗

RT @NathanMcNulty: Feeling really bad for the @Tailscale folks Lots of people have no idea how it works, that it's more than just Wireguard, the policy and ACL capabilities, that SSH isn't even on by default... It is wonderful software that someone crapped on, and now everyone wants to roll in it

评论工具 @_xpn_ source ↗

RT @reybango: Was reading a threat analysis report by @K7CyberSecurity and it motivated me to create a video about building an ISO to deliver a payload. Fun exercise leveraging techniques I learned from @_RastaMouse https://t.co/1PbbusIZWk https://t.co/MLyk6FOBO9

恶意软件 @_RastaMouse source ↗

展示 Agent 集群实际入侵链路与横向到敏感 Slack 通道的效率。

RT @lizzadwoskin: EXCLUSIVE: In less than 12 hours, a group of ex-DOGE cyber researchers were able to penetrate Anthropic’s systems using AI agents — including Slack channels where the most sensitive models are discussed. This is just weeks after Anthropic whistleblower Joseph Coxon charged the company w/launching “a hubristic gamble” from a “private company’s Slack.” Read our story here (gift link): https://t.co/x6F8Dg6SLR

ai_agent事件llm @0xocdsec source ↗

史上最大更新,口令恢复策略与 FPGA 上 bcrypt 的评估基线随之改变。

RT @hashcat: hashcat v8.0.0 released! The largest update in the history of hashcat, with PCFG attacks, hash recipes that crack hashes without a hash-mode, and bcrypt on FPGA boards. Release notes: https://t.co/0ffrO586PJ Download: https://t.co/Grn4jZmAWr https://t.co/ir6CSSsGcJ

工具提权 @pentest_swissky source ↗

10 周内近 500 次登录的完整手法链,可直接映射为检测规则。

RT @HuntressLabs: A token “spray-and-replay” campaign set the stage for business email compromise at dozens of orgs. Across nearly 500 M365 sign-ins, we saw: 📱rogue Entra devices 🧼 wiped and replaced MFA 📩 email hiding rules How tradecraft evolved over 10 weeks: https://t.co/s9abGAfiMC https://t.co/8zPCIV4NNZ

云绕过事件 @virusbtn source ↗

面向已 root 设备的固件级改写工具,附带版本相关 modem 补丁器。

mtk-imei-switcheroo — Read and write IMEI records and WiFi/Bluetooth addresses in MediaTek NVRAM files or partition images. Includes interactive ADB tools for rooted devices and separate, version-specific modem/LK patchers for biloba. https://t.co/51OrMmDCwM https://t.co/GxGmo3DIjE

工具固件移动端 @KitPloit source ↗

👀 𝗠𝘆 𝗲𝘆𝗲𝘀 𝗮𝗿𝗲 𝗽𝗹𝗮𝘆𝗶𝗻𝗴 𝘁𝗿𝗶𝗰𝗸𝘀 𝗼𝗻 𝗺𝗲 ... ​ At first glance: 𝗠𝗶𝗰𝗿𝗼𝘀𝗼𝗳𝘁 ✅ ​ Second glance: 𝗿𝗻𝗶𝗰𝗿𝗼𝗰𝗼𝗳𝘁 🤨 ​ Third glance: “Wait a minute...” 😂 ​ #BrandsImpersonation https://t.co/nX4Q2pFkhS

绕过 @0x534c source ↗

RT @anyrun_app: ⚠️ RAT families showed contrasting activity trends last week. #Romulus climbed 32% while #AsyncRAT dropped 35%, with #Remcos and #XWorm also declining. Meanwhile, #FormBook gained nearly 40%. 📌 Trend to watch: declining activity among leading RATs doesn't mean every family is slowing down. SOC teams should assess detection gaps at the family level rather than assume lower volumes mean reduced exposure. Monitor the malware driving today’s attacks: https://t.co/smcNHFuTgj #Top10Malware

报告事件 @virusbtn source ↗

演示本地 Agent 做设备取证的能力,同时暴露端侧隐私缺口。

RT @starmexxx: Claude Code found every "vanishing" photo from Instagram and every "self-destructing" message from Telegram on my iPhone. both apps promise the content disappears after you view it. neither one actually deletes it. i plugged my iPhone into my MacBook to run the same backup audit from my last post. Claude Code flagged two folders i had no idea existed. the first one was Instagram. a cache directory with every single photo and video anyone ever sent me in vanish mode. every "this will disappear after you close the chat" photo from the last 2 years. 11,400 files. all still there. all viewable. the second one was Telegram. every self-destruct photo. every secret chat image with a timer on it. every video someone sent me that said "this message will self-destruct in 10 seconds." 3,700 files going back to 2022. not one of them actually self-destructed. what was in those folders: → vanish mode photos from my ex that were supposed to disappear the second i closed the chat → self-destruct videos from a friend that had a 5 second timer on them → photos i received and never even opened that still got cached → voice messages from Telegram secret chats set to auto-delete after 30 seconds 15,100 files in total across both apps. every single one was supposed to be gone forever. the apps remove the message from your chat. your phone keeps the file. any app with file access can read it. i checked my permissions. 31 apps with Files and Folders access. weather apps. QR scanners. free PDF readers. a calculator app i downloaded in 2021. any one of them could have been looking at every "disappeared" photo i ever received this entire time. the check takes 2 minutes: step 1 → plug your iPhone into your Mac → run a backup step 2 → paste the prompt from the first reply into Claude Code step 3 → it scans every cached file that should have been deleted → shows you the total count by app step 4 → go to Settings → Privacy & Security → Files & Folders → revoke access from every app you don't use daily i went from 31 apps to 4 in two minutes. and if you ever sent a vanish mode photo to someone else → their phone cached it too. there is no disappearing message. there never was. save the prompt. check it today not tomorrow.

ai_agent移动端取证 @cyb3rops source ↗

可复用的话术素材,无新增技术细节。

Important quote ... plan to use it in your slide decks in the coming months "The ease with which open-weight AI models can be combined and spread into countless attack tools has transformed cyberattacks, which once required advanced skills, into an era where anyone can easily carry them out."

llm报告 @cyb3rops source ↗

反映攻击面扩大的舆论背景,无技术增量。

RT @AISafetyMemes: "the internet itself is on the verge of collapse" After another massive cyberattack, GMO - one of Japan's biggest brokerages - is urging the government to intervene. "The cyberspace that I built with my colleagues, believing it would bring smiles to people's faces, is now on the verge of being completely destroyed by malicious actors. The underlying issue is the explosive evolution and misuse of AI technology. The ease with which open-weight AI models can be combined and spread into countless attack tools has transformed cyberattacks, which once required advanced skills, into an era where anyone can easily carry them out."

事件llm @cyb3rops source ↗

现成的 Web 渗透测试用例集,省去自行收集。

xss-payload-list — Curated collection of cross-site scripting payloads, polyglots, and injection test rules for web application penetration testing and bug bounty hunting. https://t.co/HMQennJl3Z https://t.co/0xpVEgduQP

xss工具 @KitPloit source ↗

口令恢复主力工具跨大版本升级,自建环境可直接拉下来试。

hashcat v8.0.0 — World's fastest and most advanced password recovery utility https://t.co/BZhzMpMVxC https://t.co/mDyqtFCIdO

工具口令 @KitPloit source ↗

saw this, didn't know it had runtime code generation, thought it was just a c2 steered by ai. i had abandoned this research years ago b/c it was not quite reliable and there seemed to be little interest. i was glad to see this project and looking forward to the innovation in this area. the models are definitely capable at this point. i see no reason not to move this route. for a while, i thought, i want to be surgical to a point that i dont want to offload everything to ai, but we are about to blow past that.

ai_agentllm @Teach2Breach source ↗

统一入口简化自托管多机推理,但单点暴露全部本地模型。

RT @TeksEdge: Your home AI lab can now look like one big model server. A new open-source project, llm-router v0.1.0 puts, one OpenAI-compatible API endpoint in front of multiple local inference boxes. For example, you could have 🖥️ PC #1 → RTX 5090 → Qwen 🖥️ PC #2 → multi-GPU rig → GLM 🍎 Mac → LM Studio model 🖥️ Server → vLLM models And your apps see only → 1 API endpoint llm-router discovers which models each server is hosting and automatically sends the request to the machine holding the exact model ID. It supports the following harnesses ⚙️ LM Studio ⚙️ vLLM ⚙️ TensorFold Including these services → streaming → chat completions → Responses API → embeddings → Prometheus metrics → Docker → amd64 + arm64 MIT-licensed and written in Go. Link in ALT

工具llm自托管 @0xocdsec source ↗

man, the whole php2shell discourse is incredibly frustrating. 1. people claimed vendors were exaggerating the vulnerability and that it was practically hard to exploit. 2. they pointed to the unclaimed wallet in stealmybitcoin as "ground truth" supporting that argument. 3. except apparently nobody bothered to check how badly configured the challenge was in the first place. what annoys me is how confidently people, including security professionals, jumped on the bandwagon to declare the vulnerability overhyped based on a challenge that wasn't even properly working. like, think about whose narrative is actually more dangerous here. the vendor saying php2shell is reliably exploitable and urging people to patch, or someone pointing to a broken challenge as "ground truth" that the vulnerability is overhyped and practically hard to exploit? i agree it's hard to exploit without a frontier model or a highly skilled operator using capable open models. not everyone has access to either, and that could partly explain why we aren't seeing widespread exploitation. but does that really change how we should think about patching? you can reasonably assume well resourced threat actors have access to frontier models and talent to repro. they're not necessarily doing amateur spray and pray attacks, right? they probably have specific targets, and can afford to spend hours adapting an exploit to those environments. i don't think you're explicitly telling people not to patch. but using an unsolved, badly configured challenge as evidence that the vulnerability is overhyped can give people false confidence and lead them to delay patching.

cve事件 @S1r1u5_ source ↗

RT @S1r1u5_: weren’t you the one claiming that the unclaimed wallet was finally giving us “ground truth” on how hard php2shell is to exploit, and calling other claims bullshit? did you consider that the challenge itself was broken for the first ~10 hours, and then the single-threaded server got DoSed? how is any of that ground truth on real-world exploitability? i said the challenge could be solved in under 5 hours once they fixed the setup.

cve事件 @S1r1u5_ source ↗

RT @dealignai: Nanbeige 4.3 33B FULLY UNCENSORED GGUF The new Qwen 27B competitor! Nanbeige's own benches have the base model ahead of Qwen 27B, 512k context. Compliance for thinking off and on, 200/200. Every category at 100% (chem/bio, cyber, illegal, misinfo, harassment, harmful), zero loops. MMLU general a/b/c/d has ~2% parity with the base quants, but the smallest iq2xss shows some obvious damage. Each quant was calibrated with my own corpus and have imatrix. Q6 is 25gb, Q4 is 18.7gb, IQ2 is 8.7gb. https://t.co/nA6BoGMs5w

llm自托管 @0xocdsec source ↗

RT @ItsmeAjayKV: My unlocked RTX 3090 is really stable. Been continuously running Qwen3.8-27B for last 14hrs 🤩 Since I'm running everything that pushes GPU to extreme am not going to turn this off. Next tests: - video generation with MiniMax H3 - Image generation with Qwen-Image-2.1 and Ming-Image. Reply if there is anything else do you me want test.

llm @0xocdsec source ↗

RT @SemiAnalysis_: PROFIT ALERT🚨: AMD @vllm_project can generate up to $3/hr/chip in pure profit serving the open-source Kimi K3 model, even after accounting for the rapid increase in chip rental prices post-Claude Code. Kimi K3 2.8T is a great model for AMD, as it has over 2.8 trillion parameters that can take advantage of AMD's 288GB per chip. B200 CANNOT fit Kimi K3 in a single server, so it has to use tensor parallelism or pipeline parallelism for communication across 400Gbit/s InfiniBand/Ethernet, which is 18x slower than the BW within an NVLink domain.

llm云 @0xocdsec source ↗