⚡Megatron LLM Hub
All digests
Twitter

推特安全流

Top7

自托管 Agent 直接做主机侧检测与处置,是本地 AI Agent 安全落地的现成样本

kodeblue — SOC, incident response and SecOps for Linux servers in one agent: learns each host, watches it with zero-token sensors, triages with a small model, and contains attacks under a root-owned policy. Optional hub for fleets. https://t.co/qicOWXId4v https://t.co/QZn6FJjLBx

ai_agent工具事件响应 @KitPloit source ↗

走每台 DC 默认开放的 9389 端口,RBCD、DCSync 无需另起服务即可执行。

RT @g0h4n_0: Releasing a new tool « adshell-rs » an interactive Active Directory attack shell over LDAP(S) or ADWS (SOAP) 🔥🦀 RBCD, Shadow Credentials, DCSync, machine accounts, password reset, read objects… through the SOAP door (tcp/9389) every DC exposes by default 👇🏼 #ActiveDirectory #ADWS #RedTeam #Rust #Pentest

工具ad红队 @0xocdsec source ↗

AI 代理框架已被实际武器化,用于扫描、匿名上传与 SSRF 利用并投递开源 RAT

RT @MsftSecIntel: Threat actors’ abuse of ARTEX continues to expand, with Microsoft Threat Intelligence observing financially motivated Storm-3358 using the agentic AI offensive framework alongside conventional automated tools to target financial organizations. https://t.co/SJn0rBvsU7 Storm-3358 has used ARTEX to scan internet-exposed systems, exploit security flaws including anonymous file upload and server-side request forgery (SSRF), and install open-source remote access trojans (RATs) for persistence and lateral movement. This activity illustrates a continuing trend of threat actors using agentic AI penetration testing frameworks to accelerate attack chains. Read our latest blog to get indicators of compromise (IOCs), along with mitigation, detection, and hunting guidance.

ai_agentllmssrf事件 @artem_i_baranov source ↗
Must-see13

证书加 TLS 指纹的溯源方法可直接复用于自有基础设施扩线

RT @MichalKoczwara: Just finished the full walkthrough of my TraderTraitor infrastructure research today. From one reported ROOFDECK C2 to 13 linked endpoints using certificates, TLS fingerprints and historical records. Full walkthrough. Hope you enjoy it, and happy hunting! https://t.co/aKwxOeFUiE

报告威胁情报证书 @domchell source ↗

利用原语与复现细节齐全,SELinux 强制与锁 bl 下仍可行

RT @0xManan: Android LPE to uid-0 + KernelSU from untrusted app (SELinux enforcing, locked bootloader) CVE-2026-98260 · posix-cpu-timers UAF after failed non-leader execve - de_thread() leaves timerqueue node; exit_itimers() frees k_itimer; cleanup_timerqueue() rb_erase() on freed object - AF_UNIX spray reclaim → 8-byte write primitive → fake/real cred - Verified on Honor MagicPad3 Pro (kernel 6.12.38 / 6.12.58) + other ports https://t.co/GbutJe2LaH #Android #kernel #LPE #CVE

lpe内核cve移动端 @0xocdsec source ↗

为本地 Agent 提供可抛弃隔离执行环境,直接缩小跑不可信代码的爆炸半径。

RT @shanselman: A lot of people don't realize that Windows has a built-in feature called Windows "Sandbox" that boots up a brand new virtual Windows machine in seconds. So I made an MCP server so you can test your apps inside a throwaway machine in seconds. https://t.co/huB9hAoxSv

mcpai_agent隔离工具 @kernullist source ↗

给出不需专用网络安全模型、低成本复现漏洞挖掘的可操作路径。

some alpha drop for doing inexpensive vuln research. instead of writing an article on the prompts/process behind this bounty (and more), I am dropping some prompts and the thought process I used as part of this research. ik you might say prompts don't matter much, frontier models' baseline intelligence can rawdog vuln discovery (and it is true to a good extent), but there is a high chance that you either do not have access to cyber models, or even if you do, they are expensive (and with heavy guardrails). The idea is, these prompts work well with open-weight/less expensive models. Not saying these will outperform the raw intelligence and capabilities of frontier models, but open-weight models perform much better with a bounded/goal-specific approach (where the goal is set, and the search space is massive) rather than open-ended exploration (where the goal is vague, and the search space is massive). anyway, here are some prompts. I don't need to share much, because you can just create variations of these and apply them wherever you wish. - find places where the app is supposed to enforce something, but doesn't actually enforce it. - find security rules that are enforced in one part of the app, but just assumed to be true in another. - find security assumptions the app relies on, but never actually checks or enforces. - find places where the app depends on undocumented behavior, and what happens if that behavior changes. - find operations that are secure on their own, but become vulnerable when you combine or chain them together. - find cases where the same object can be accessed in different ways, but each way ends up having different permissions. - find objects whose identity can be changed or rebound to another user, without going through the right checks. - find requests that get rejected or fail, but still end up changing something or triggering a side effect. - find security checks that only work if certain events happen in a specific order, and what happens if that order changes. - find operations where the transaction fails or rolls back, but some of the side effects still happen. - find things you can still do with an object even after it has been deleted, or access to it has been revoked. - find state changes that can be reversed or undone by someone who shouldn't have permission to do so. - find objects that can end up in two contradictory states at the same time, and what operations become possible because of that. I take such prompts, put it in a mutator (which has threat model of the project I am doing sec engagement), the mutator will mutate the prompts based on the threat model, and spawn swarm of agents, with different starting prompt and a bounded goal, the process repeats till all paths are exhauted, or some new signal is generated, which then feeds back to mutator. Doing this using a cheap model is very effective and sometimes it might yield you something even expensive models miss, we are just trying to maxximise the search space, from as many angles as possible, and once this process generates enough signals, just use a frontier model to confirm/triage issues, or for finding similar variants of such issues, that's basically it. There is v less moat is making a custom harness (unless you are an elite vuln research who can codify their approach in it or a 100x agentic engineer, I am neither).

llmai_agent方法论 @0xAsm0d3us source ↗

利用尚未扩散,但已加速;暴露在互联网的 SSH 服务自查窗口正在收窄

🚨During the past 24 hours, we have observed a small amount of actors targetting SSHole (CVE-2026-123158) Exploitation is not widespread yet, but is showing an early accelerating trend. Track exploit intelligence 👉https://t.co/FMlpPAiZfm https://t.co/fhCpCN3iFk

cve事件利用 @DefusedCyber source ↗

每个 DC 默认开放的 9389 端口即可发起多类域攻击,红队上手成本低

RT @g0h4n_0: Releasing a new tool « adshell-rs » an interactive Active Directory attack shell over LDAP(S) or ADWS (SOAP) 🔥🦀 RBCD, Shadow Credentials, DCSync, machine accounts, password reset, read objects… through the SOAP door (tcp/9389) every DC exposes by default 👇🏼 #ActiveDirectory #ADWS #RedTeam #Rust #Pentest

工具提权active_directory @Jean_Maes_1994 source ↗
Recommended15

GPU 驱动侧的内核 r/w 原语,为后续 SPTM 绕过研究提供起点

RT @0xjohnny: And now: iOS AGX r/w via ClearSword krw. No, this is not an SPTM bypass. It may help further research, but on its own it cannot access SPTM-protected memory without faulting. https://t.co/V4eP1lv9OQ

移动端内核poc @0xocdsec source ↗

AI 智能体批量复现真实漏洞的能力指标,反映攻防自动化的当前水位

RT @NFT_Chen: 💥没想到!360自研AI攻防系统图龙锋在AI安全世界杯CyberGym以99.4%漏洞复现成功率登顶全球第一,碾压GPT-5.5-Cyber、Mythos Preview、Gemini 3.8 Flash Cyber! 🔹三轮评测平均99.38%,挑战1507个真实漏洞 🔹远超此前最佳组合11.9% 🔹安全大模型+智能体蜂群技术路线 🔹已挖漏洞超1.5万个,全国唯一通过信通院评估 #360图龙锋 #CyberGym #AI安全 #周鸿祎 #GPT55Cyber

ai_agentllm报告 @0xocdsec source ↗

开源 C2 框架新增图形会话直连,内网横向操作门槛进一步降低

Hopefully coming to Mythic v4 soon. Direct RDP and VNC functionality via Guacamole. No rportfwd involved, just a direct feed to Guacamole via a new interactive streaming mode from the mythic agent. Might be a tough watch from a mobile device, so I'd recommend a larger screen. https://t.co/l6ATcfP5nd

工具rdp开源 @_EthicalChaos_ source ↗

指出 AI 代理规模化攻击下“检测到也难以处置”的现实缺口。

RT @MTSlive: Calif CEO @XorNinja says cybersecurity has no playbook for stopping 1,000 AI agents attacking a network simultaneously, even if defenders can detect them: "Attackers always adapt faster and earlier than defenders. Even the AI companies themselves don't have great security." "Imagine you enter a cage fight, but you are blindfolded and the other guy stay in the dark, and every once in a while they would come out and punch you and then disappear." "How could you win that fight?" "I was at Google and that's basically the mantra at Google. In security you have three things: prevention, detection, and response. And Google accept that prevention gonna fail at scale." "You can detect them, but does it help you if you find that there's maybe a thousand agents hacking you right now inside your network? How can you contain them without a full shutdown of your system?" "This actually requires defense to come up with a new paradigm. I haven't seen anybody come up with a good idea here. We're not there yet." @calif_io

ai_agent自动化观点 @lcfr_hax source ↗

云端到车控的整链细节可复现,是车联网攻击面分析的完整范例。

RT @infosec_au: Earlier this year, I hacked a major automotive company with @samwcyo. We were able to remotely unlock cars in the AU and US. My writeup for AU is here: https://t.co/UAjPrav4JC and our writeup for the US here: https://t.co/4FzM1ia736

车联网云报告 @0xocdsec source ↗

AI 重写核心组件时,用差分对比和 fuzz 做回归验证,是发现语义漏洞的可行手段

RT @zacker150: Actual software engineer here. Vibe coding is when you use AI to write software without manually reviewing every single line of code. Instead, you rely on empirical tests, formal verification, and static analysis to verify that your code is correct. Rewrites like the Discord rewrite are prime targets for vibe coding because you can easily verify them. Just take the old code, run it with a ton of random inputs (this is called fuzzing), and compare the output with the new code. Your rewrite is done when the output is always the same. As for how you know it's vibe coded? Everything is vibe coded these days. Software engineers don't write code by hand anymore. Instead, we herd clankers writing code.

llm代码审计方法论 @Teach2Breach source ↗

自建云环境可以一键运行配置审计基线,成本低于商用工具

FuFuFaFa — Framework for Unified Flaw & Fault Auditing The ultimate AWS Cloud Security Audit Framework that makes security auditing so easy, you can do it even when you're sleepy! https://t.co/njLAt1HcI5 https://t.co/4X43xI9ebH

工具云审计 @KitPloit source ↗
More47

厘清 AI Agent 攻击的效能边界,有助于确定检测侧重点

RT @HackingLZ: So much “wait until the agent swarms are stealthy” talk. They’re effective today because they do a massive amount of noisy testing and iterate on the results. Making them stealthier means sacrificing coverage. That tradeoff has always existed. This is the same pentest vs. red team conversation we’ve been having for years. You get coverage or stealth you don't get both.

ai_agent红队议题 @0xTriboulet source ↗

调试 V8 漏洞与 coredump 时可直接看到 JS 实参,省去手工翻栈

RT @JoyeeCheung: Landed JS argument inspection support in the GDB & LLDB plugins for V8, for both live and coredump debugging - bt now shows previews of JS arguments for JS frames - New `v8 args` command to show arguments of selected frame - `v8 inspect` now takes this, arg names and indices https://t.co/xp8wCwvLnW

浏览器工具v8 @0xocdsec source ↗

系统完整性校验成了准入门槛,自定义 ROM 用户直接失去支付可用性

RT @Dennis_Rye: Europe built Wero so we wouldn't need Visa, Mastercard and PayPal… and now the Dutch bank ASN only lets you activate it if Google approves your Android phone. The app runs Google's Play Integrity check, so anyone on GrapheneOS or /e/OS is locked out. The bank's answer: Wero relies on the "standard security controls for regular operating systems". So the European payment system meant to get us out of US dependence gets Google as the gatekeeper… sovereignty with an asterisk

移动端绕过android @vanhoefm source ↗

RT @QwenAIFans: ⚡️ DFlash2 just took Qwen3.8-27B on one 3090 from 201 to 296 tok/s. Same INT4 prompt, greedy. Four users at once still hit ~401 tok/s total. Caveat: author 3090 run; this engine is 3090-only so far. via tuan3w / qwen-warp 👇 https://t.co/O9yi1ngNpj https://t.co/ld44O3Hj2z

llm自托管 @0xocdsec source ↗

THIS! but go an extra mile and hook it up with truth oracles (search APIs) for real time truth. The actual truth comes from real sources, AI just helps orchestrate, summarise, and produce final output. Many of my deep research agents works on this simple system. A final frontier model is used for checking correctness/completeness tho, and that's it. This is very inexpensive.

ai_agentllm @0xAsm0d3us source ↗

I genuinely don't think people in India understand how quickly AI research is progressing in the US right now, or how enormous the gap is becoming... All this happened in last few days: - Thinking Machines (Nov 5): Inkling 2 Small, 594B MoE, 17M+ RL rollouts, 3T+ generated tokens, asynchronous RL, NVFP4 training. - Ricursive (Nov 4): Ripple-65B, 12T pretraining tokens, 1M context, 80.5 on TerminalBench 2.1 using just 1.4M RL rollouts. - Cognition (Nov 4): SWE-2 Flash, trillions of mid-training tokens, specialized coding RL, 3x faster inference at 5x lower cost. - OpenAI (Nov 5): GPT-6 Cyber, 48.5% on ExploitGym, specialized security post-training, autonomous vulnerability research. ik we are already doing a lot things to catch up, but we really to pace tf up and probably reevaluate our approach.

llm @0xAsm0d3us source ↗

路由器只买二手 路由器是纯固态系统,全机零机械运动部件;且以读为主,系统开机解压后常驻内存,动态转发全在 DRAM 中以电荷生灭,底层闪存几乎常年只读不写。只要散热正常,核心硅基芯片的物理退化周期远在 10 年以上。 既然物理性能毫无损耗,买二手就是拿极致性价比换取 100% 的转发能力

硬件 @nodotbtree source ↗

表明 Agent 已能辅助产出密码分析级新结果,超出常规漏洞扫描范畴。

RT @qwerty__cs: Google DeepMind's AI agents found two new weaknesses in legacy 1024-bit RSA, in just a few weeks. Working under human guidance with cryptanalysis experts from UC San Diego and Inria, the agents came up with: • A signature forgery method that, when the attacker controls only a small part of the signed data such as a 32-byte digest, is in theory about 1,600x faster than the previous best for 1024-bit keys • A batch factoring method that cuts the cost of each additional 1024-bit key by about 30x compared with earlier batch techniques RSA-2048 and larger keys aren't affected, so banking, messaging and software updates are fine. The ones exposed are older systems still running 1024-bit keys. https://t.co/MqxvWXGtuU

ai_agentcrypto报告 @0xocdsec source ↗

移动端固件级后门风险,采购二手或第三方定制机时需核对来源与签名。

RT @GrapheneOS: @mano_crypto These phone sellers usually sell forks of GrapheneOS with a remote management system to force people to keep paying for the device. Many of these devices are likely deliberately backdoored. It's proven the police in these countries conflate these devices with actual GrapheneOS.

移动端固件后门 @0xocdsec source ↗

含具体漏洞成因与利用路径,是成体系拆解漏洞研究的少见材料。

RT @CodeanIO: We released our latest blogpost covering the details of some bugs that @b0n0b0__ presented during Romhack Camp (hosted by @cybersaiyanIT). They'll publish the video of the talk and the slides, make sure to check all that out and stay tuned for part two! https://t.co/5hU9JIKyDD

漏洞报告poc @0xocdsec source ↗

i'll post more about my system for spinning up such swarms of agents with bounded goals and how everything comes together with a memory + RAG (it has some specialized writeups, niche info, this is mostly for blackbox stuff though) soonish, but you could just vibecode something similar by studying the architecture of deepseek harness (especially it's plugin, subagents architecture) / omp / opencode etc repos. These repos are already v good, but for my usecase, I mostly need something small which allows me to do as much customizations as possible, so making a toy project made sense.

ai_agentllm @0xAsm0d3us source ↗

RT @g0njxa: Another party in Moscow 🇷🇺 sponsored by cybercrime ZELENKA / LOLZ forum, featuring their LOLZTEAM marketplace for stolen accounts. A party for Halloween 🎃 Allegedly “attended by over 5,000 people” and with famous rappers like “Fortuna 812” on stage And allegedly a safe place to meet between threat actors, discuss about gray and black activities and advice newcomers into the criminal field

威胁情报事件 @0xocdsec source ↗

若属实则绕过 MIE 硬件缓解,打破近一年无公开 iOS 利用的局面

RT @MTSlive: Calif CEO @XorNinja reveals how his team hacked an iPhone and maintained root access even after rebooting or updating the device, challenging one of the most common pieces of iPhone security advice: "In software security, the iPhone is like the holy grail of vulnerability exploitation." "Since Apple introduced MIE, which is a hardware mitigation mechanism, since last year, nobody actually has published any public iPhone exploit. So we really, really wanted to hack the iPhone for that reason." "We managed to get root access to the iPhone and even persistent access." "In the past people said one of the best practices for iPhone security is to reboot your phone after a while because all the malware would go away." "But we show that there's ways to persist our attacks even if you reboot or you upgrade your iPhone." @calif_io

移动端提权绕过 @0xocdsec source ↗

依赖 Claude API 的自动化与代理流程需要预备账号随时失效的替代方案

RT @Appraiser008: Claude 现在基本把国内用户的路堵死了。 不管是刚打算注册的新账号,还是已经续费在用的 Pro 老号,随时会弹出强制 KYC。 新号很难绕开这一关,老账号也不再是保险箱。 而中国大陆不在它支持服务区域,就算手里有证件,核验这一步本身就是一道很难跨过去的门槛,一旦触发验证,账号随时受限甚至失效。 之前那种靠环境苟住的玩法,风险一下子拉高很多。

llm风控账号 @0xocdsec source ↗

说明 AI 代码审查发现的仍是 2001 年级别的老问题,评审流程缺位才是根因

RT @Laughing_Mantis: character encoding issues is literally what worms and IIS exploits used in 2001 used. Not even double encoding, which I wouldn't be surprised if that was a second bug. Secure web development practices and web scanners from the early 2000s both would have caught this. Not sure what's more embarrassing not finding this in development or bragging that an AI caught this

绕过web议题 @daaximus source ↗

RT @QwenAIFans: 🔥 One 3090 just beat Strata on Qwen3.8-Flash-Next. llamAmpere v0.6.3: 120+ tok/s, >3K/s prompt. 12% faster than Strata, 27% lower error than Unsloth Dynamic v3.1 3XXS. Caveat: author's AmpereInteger+ bench on his 3090+64GB. via @JakeKAllDay 👇 https://t.co/idin9xLOQb https://t.co/bLoBAt8Vqh

llm本地部署性能 @0xocdsec source ↗

依赖 JIT 开关等平台限制绕过,移动端跑桌面系统的权限路径值得留意

RT @qwerty__cs: macOS is now running on an iPhone. Henge 0.1 just came out. It downloads Apple's own restore image and installs macOS Sequoia right on the phone, no Mac needed. On an iPhone 17 Pro it reaches the lock screen in about 15 seconds, and Safari, Chrome, VS Code and Discord all work inside it. No jailbreak needed, but you do need: • iOS or iPadOS 17 or later • A14 or M1 and newer • About 43 GB free • JIT enabled (StikDebug, TrollStore, a jailbreak, or the built-in StikJIT helper on iOS 26+) Sideload Henge.ipa with SideStore or AltStore, or install Henge.tipa with TrollStore. https://t.co/VO3Hsc8OMy

移动端工具绕过 @0xocdsec source ↗