⚡Megatron LLM Hub
All digests
Twitter

推特安全流

Top7

只升级 WordPress、Drupal、Joomla 不够,必须升级 PHP 本体才能修复。

RT @nullhuntersec: PHP2Shell: vulnerabilidades críticas en PHP permiten ejecución remota de código sin autenticación. PHP ya publicó el parche. Actualizar WordPress, Drupal o Joomla no alcanza: hay que actualizar PHP. 🔗 Referencias en el hilo. #PHP2Shell #Ciberseguridad #AppSec https://t.co/hnTG4VWGnP

rcephpcve供应链 @shadkz_ source ↗

环境特定利用已出现,且几乎全由开源权重模型生成,落地门槛显著降低。

RT @RevToJa: Universal exploit for #php2shell is rather impossible, but specialized exploits for known environments are coming. I tried my luck with pfSense 2.7.2 - PHP 8.2.11, FPM, 512M limit. Variant with session file write to a crafted path. Almost entirely written by an open-weights LLM. https://t.co/WUEiZDMBEE

rcepocllmphp2shell @0xocdsec source ↗

十月已修补但 PoC 公开,域内默认防护失效,需核对补丁状态。

RT @GrayHatKiller: 1/2 LDAP signing was supposed to stop NTLM relay to LDAP. I found a way to relay coerced SMB auth to a DC with LDAP signing and channel binding enforced. Patched in October. Blog: https://t.co/dUHIN6JfaE PoC: https://t.co/xT98z64V96

ad中继绕过 @_xpn_ source ↗

本地即可跑,把提示注入与 Agent 越权测试接进 CI/CD 流水线。

promptfoo v0.125.0 — Test your prompts, agents, and RAGs. Red teaming/pentesting/vulnerability scanning for AI. Compare performance of GPT, Claude, Gemini, DeepSeek, and more. Simple declarative configs with command line and CI/CD integration... https://t.co/WV0HP86XrD https://t.co/Vg1YRECq1N

ai_agentllm工具红队 @KitPloit source ↗

已有公开 PoC 且 10 月修补,域内提权路径可复现。

RT @GrayHatKiller: 1/2 LDAP signing was supposed to stop NTLM relay to LDAP. I found a way to relay coerced SMB auth to a DC with LDAP signing and channel binding enforced. Patched in October. Blog: https://t.co/dUHIN6JfaE PoC: https://t.co/xT98z64V96

中继提权pocad @_EthicalChaos_ source ↗

可用自然语言驱动搭 fuzzer,明显降低自建模糊测试的起步成本。

RT @secorizon: We just released OffByWon 2.0, an advanced network/binary fuzzer framework. Lots of changes, more powerful, easier to implement. We also built an AI module, so you just need to instruct your AI to build a fuzzer using OffByWon, the framework will do the rest of the magic :) https://t.co/nfCUaCxkk1

fuzzingai_agent工具 @0xocdsec source ↗
Must-see17

硬件加密产品的信任模型直接失效,同型号设备里的数据等同明文。

RT @IntCyberDigest: ‼️ An encrypted USB drive from SecZure has a backdoor that opens its files without the user's password. Two built-in accounts with fixed passwords decrypt them, Binary Ninja's Xusheng Li found. OpenAI's Codex broke Netac's FileLock in 7 minutes and saw it logging passwords. https://t.co/btN5WylDRz

固件后门逆向事件 @0xocdsec source ↗

给本地 Agent 接逆向、取证工具,能明显缩短根因定位的路径。

RT @rae1101x: your claude code is debugging blind. native crash. no logs. no source code. no debug symbols. 4 turns to find the cause. give it the ability to reverse-engineer the binary? 1 turn. you can’t prompt your way into evidence that isn’t there. https://t.co/Bn1JpnmI49

ai_agent逆向工具 @morluto source ↗

列出了文件、网络、MCP、凭据的隔离边界,可直接对照自建 Agent 方案。

I am evaluating secure sandbox surfaces for high-trust agent environments (benchmarking, SFT, expert on-policy supervision, outcome based RL, etc). Obviously there are some pre-baked solutions, one that is attractive is Docker's sbx (https://t.co/e661p2xppI). Sbx runs agents inside isolated Linux microVM's with their own kernel and private Docker Engine. They use filesystem passthroughs and host-side bridges to facilitate file access, networking, mcp tool usage, credentials, etc. These microVM's offer a very strong execution boundary but obviously the brokers are a major attack surface. Looking through the changelog for sbx, the situation is a bit more grim than you would expect. Between June 1 and today (Oct 24) there are 9 CVE's including critical issues (like a macOS virtio-fs escape CVE-2026-77179). These look like typical logic issues you might expect. What's interesting though is that the changelogs also reveal areas that got security fixes but no CVE. For example, a full broker mediated guest-to-host escape, shown in the poc below (on v0.39.0 before Sept 7). Some light reverse engineering was required since you don't get core implementation source. Pre-patch a sandbox could supply DBUS_SESSION_BUS_ADDRESS through the guest-controlled host_env sent to the host browser bridge. On the Linux host, the bridge passed that environment to xdg-open, whose desktop-detection path invoked dbus-send, which accepts a "unixexec:" address leading to code execution.

ai_agent隔离沙箱 @FuzzySec source ↗

覆盖 mod_php 与 php-fpm,主流 CMS 默认部署即受影响。

RT @dguido: Update: Our PHP exploit is now production-grade. * stock php.ini-production, display_errors=off, normal memory_limit, multi-worker, ASLR, under load * ARM64 + x86-64 targets * MediaWiki, WordPress, phpBB, Laravel * Apache/mod_php and nginx/php-fpm -> root https://t.co/vXQjheyp20 https://t.co/CAF1htNW0o

rce提权pocweb @artem_i_baranov source ↗

迄今少见的 agentic AI 加速实战后渗透的公开观测,附狩猎查询。

RT @Unit42_Intel: An update to our threat brief on PHP CVE-2026-109618 & CVE-2026-109619: We observed an attack using agentic AI platforms for rapid post-exploitation and access to MFA secrets within 2 hours. Read the updated brief for analysis and hunting queries: https://t.co/QFY5mbrGpx https://t.co/dCCEKRbc4M

ai_agentllm事件报告 @cyb3rops source ↗

攻击门槛从写脚本降到对话,防护假设需整体前移

i think we should just stop arguing whether a layman with no security experience can do damage. they absolutely can with llms, basic computer knowledge and knowing the language. there is no point arguing about this anymore, its obvious what a model like gpt-5.6 sol can do. also this is not some old script kiddie exploit spray and prays, the model adapts and improves per target. just assume it is like handing someone a gun that aims itself, point and shoot, like, just stop asking whether they have the skill to pull the trigger. the only constraints left are literally law and order, the fear of committing a crime, and maybe morals.

ai_agentllmrce @S1r1u5_ source ↗

AI 助手可直接改审计策略,MCP 侧的权限与审批边界值得先评估。

RT @magicswordio: Audit event review has ruined more Friday afternoons than we can count, so we gave it to your AI assistant. MagicSword MCP: 21 tools for Claude, Codex and Cursor that pull the events, draft the policy change and then wait for your approval. Check it out on our refreshed site: https://t.co/KvSSji8GVD

ai_agentllm工具 @nas_bench source ↗
Recommended15

给出是否、何时打补丁的判断依据,避免盲目升级或漏补。

based on the ongoing discourse and conversations with other researchers, i think it would be reasonable to communicate the following points to affected software owners so they can make informed decisions about patching php2shell. https://t.co/34FYfJOfxt

rcephp2shell报告 @S1r1u5_ source ↗

给出可对照的利用前置条件与阻碍因素,用于评估实际暴露面。

RT @CrowdStrike: CrowdStrike Intelligence has reproduced variants of remote code execution (RCE) tied to PHP2Shell. Why hasn't widespread exploitation been confirmed? Our researchers explain the technical obstacles and what organizations should know: https://t.co/pgcxzsTgFa https://t.co/5xwfKC9YA4

rcephp2shell报告 @artem_i_baranov source ↗

无需原生 GSSAPI 或 ldap3 分支,Python 里即可做签名与通道绑定的 LDAP 操作。

Adds protected Active Directory binds to synchronous ldap3 connections. It uses: ✅ Impacket for Kerberos and NTLM authentication and leaves LDAP queries, entries, controls, and binary values with ldap3. ✅ It does not require a native GSSAPI library or an ldap3 fork. https://t.co/tUqbZ9gpZ5

工具adldap @ipurple source ↗

显示从渲染进程一路打通到虚拟机逃逸的完整 0day 链仍可批量构造。

RT @uz56764: We just received a $250,000 Chrome Full Chain bonus! 🎉 We've claimed the second Full Chain bonus of 2026. Only two remain! Under the rules, every vulnerability had to be a non-duplicate 0-day, and the entire exploit chain had to be completed before any of them were patched. Huge thanks to @nebusecurity for discovering a V8 vulnerability in just a single day! Also, a big thank you to @_deayzl for collaborating with me on the V8 Heap Sandbox research. P.S. KVM escapes have been getting a lot of attention. I actually captured the Full VM Escape flag in KVMCTF about three months ago. Hopefully, I'll be able to share that research soon, too!

浏览器v8逃逸事件 @FuzzySec source ↗

启动早期植入可绕过常规内核与用户态检测,落地细节值得跟。

RT @ISMGEvents: Your OS hasn't loaded, but the attack has already begun. 💀 At #NullconBerlin2026, Ido Veltzman reveals how a Linux UEFI bootkit can deploy a rootkit, complete with a live demo! 🎥 Hear from @idov31 himself on what went into the research! Register now: https://t.co/UWb3Whm1Il https://t.co/LGRB4bv5Z1

固件uefirootkit议题 @Idov31 source ↗

与厂商结论互为参照,可用于校准自身暴露面的处置优先级。

RT @JohnHultquist: A brief update on the critical vulnerabilities in PHP:  We have not yet validated exploitation in the wild and we have no evidence of successful wide scale exploitation. There are currently several exploit PoCs publicly circulating, many of which appear to be "vibe-coded" and are incapable of remote code execution. We have validated some PoCs, however, that are capable of RCE against default deployments of mod_php but not PHP-FPM. Nonetheless, exploitation may have already begun in a limited fashion and the threat is viable and very serious. Defenders should: ​1) Patch & Restart by upgrading to PHP 8.5.12, 8.4.27, 8.3.36, or 8.2.35 and restarting web services ​2) Layer Temporary Safeguards at the network edge (WAF/proxy rules) and server configuration level if immediate patching requires a maintenance window ​3) Protect Credentials by monitoring for abnormal session traffic and rotating keys and passwords on compromised internet-facing PHP systems exposed prior to patching 🔗 We are continually updating our Cloud Security blog post on this topic with real-world guidance and actionable defense strategies as we learn more. Please keep monitoring these posts for updates as you prioritize your patching efforts: https://t.co/uDnd6gk988  🔗 Technical details from Wiz: https://t.co/RqmTqEDM2V 🔗 Details of the exposure from VirusTotal: https://t.co/Toix3AM0yq

rcepoccve @artem_i_baranov source ↗

直接反驳“难以利用”说法,并给出具体受影响组件。

RT @dguido: Re: the widely circulated blog claiming the PHP bugs aren't easily exploitable: @trailofbits has a working exploit that gets a shell on an unmodified MediaWiki installation. https://t.co/mSr6Vkc4pj https://t.co/WLFLGj9i9V

rcepocweb @artem_i_baranov source ↗

免真机做 iOS 固件补丁与越狱变体测试,降低研究门槛

vphone-cli v2.12.0 — Boot and manage virtual iPhones on Apple Silicon with firmware patching, jailbreak variants, and security research features for iOS testing and analysis. https://t.co/cvBxKjvRlG https://t.co/4OvV5SyQKC

工具移动端固件 @KitPloit source ↗

LLM 与现成利用链结合后的落地程度,是评估风险的关键变量

right now, http2shell or php2shell can be given to an llm and improvised per target, point and shoot. but we don’t see mass exploitation, we dont see bad guys causing damage, why? is it because open models are not there yet (i haven’t seen any open model doing it faster yet with considerable hand holding unlike gpt-5.6 sol), or because real life has constraints like committing crime, law, getting caught? we don’t know. all we see is good guys finding and patching? so maybe the answer is to proliferate and let more good guys do this? or are we all overestimating the risk?

llmrce报告 @S1r1u5_ source ↗

模板库更新直接决定漏扫覆盖面,自建扫描应跟进校对

nuclei-templates v10.5.1 — Community curated list of templates for the nuclei engine to find security vulnerabilities. https://t.co/P61TaRAOMD https://t.co/0rIUO02vz0

工具poc @KitPloit source ↗

把 LLM 判断接进 fuzzing 循环,并讨论取舍而非只做演示。

RT @OffensiveAIcon: New talk up OAIC 2026: Vibe Fuzzing with Sam Lovejoy @_dinolover38 of IBM X-Force Offensive Research. Sam explores how graph-based inputs, program state, and LLM guidance come together in an AI-guided fuzzing engine- and the tradeoffs behind it. Watch the full talk: https://t.co/zFcLR5AvZd #OAIC #OffensiveAI #Fuzzing

fuzzingllm议题 @bohops source ↗

免去手工整理,能快速搭出 LOLBin 相关 PoC 与检测规则。

@whokilleddb Sounds like a fun project! I would recommend checking out the API endpoint on the LOLBAS site. It may help you save some POC time for building something bigger: https://t.co/Wj7hwxK584

工具绕过检测 @bohops source ↗
More79

无人机攻击面从近场推进到远程可打,漏洞细节未公开但方向明确。

RT @lyq_sqsp: At today’s GEEKCON, a researcher demonstrated what I believe is the first remote full-chain exploit of a latest WELL-KNOWN drone. Our Pwn2Own-style live competition never released vulnerability details on those success trials. Actually we did not even show the vendor’s name in plaintext. However, a guy from that vendor called POLICE on-site to take the researcher out for “investigation”. It’s extremely disappointing that in 2026, some still view security work from such a bad perspective.

exploit固件事件 @0xocdsec source ↗

按 KEV 与凭证、配置链排序修复,比按漏洞计数分配资源更贴实际。

RT @snehalantani: This take completely misses the plot. It’s never been about the number of CVE’s. Being vulnerable doesn’t mean you’re exploitable, most CVE’s are not exploitable. Most cyber attacks DONT use CVE’s, it’s about chaining together compromised credentials and misconfigs. Very few Anthropic findings became CISA KEV’s, that’s the metric that should be tracked

cve风险管理 @0xTriboulet source ↗

窃听装置伪装成常见物联网模块,办公物理安全检查可直接借鉴。

RT @spycraft101: Ukrainian counterintelligence personnel removed a hidden Russian camera and microphone system in March 2026 from the office of the chief engineer of TechEx, a major drone manufacturer. Analysis by Darren Kitchen of Hak5 determined it was an inexpensive IP camera module board with audio capability mounted behind a thermostat on the wall of an office. According to reporting from 24tv, the SBU intelligence agency discovered the camera earlier and informed TechEx but left it in place in order to pass false information back to Russian intelligence. TechEx manufactures the Striker Mini interceptor and the Stalker line of fiber-optic drones. More than 70 combat units in Ukraine are currently equipped with TechEx drones. TechEx made a public announcement of the camera and the accompanying disinformation operation on their Facebook page on March 24th. #ukraine #russia #drone #camera #spy

硬件窃听事件 @0xocdsec source ↗

漏洞不可能被穷尽修补,检测与响应必须与补丁并行投入。

RT @spendergrsec: "I No Longer Think We’re Gonna Be Able to Just Patch All the Bugs", what I've been saying since the beginning :)

补丁观点 @S1r1u5_ source ↗

旧卡自建推理的成本路径,可用于评估本地部署的可行性。

RT @nullhuntersec: Se pueden correr LLMs en una Tesla P40? Exploramos el tema en un nuevo artículo de Null Hunter, escrito por uno de los integrantes del equipo. Leelo acá 👇 https://t.co/Xf2vcNGoCg #LLM #AI

llm自托管工具 @shadkz_ source ↗

从 XSS 到命令执行的完整链条可复现,WordPress 站点需重估补丁节奏。

RT @pwn_ai: It looks like they figured out the end to end PoC from our August Wordpress RCE chain, XSS2Shell. Our original research describes how we discovered it: https://t.co/Fql6aLPrtT

rcexsspoc @0xocdsec source ↗

WebDAV 成为杀软盲区,可被用于投递载荷与规避检测。

RT @MSNightmare2000: As of now the latest version of windows defender is unable to scan webdav shares because of the way how it open files. Any attempt to read from a webdav made by defender will fail no matter what.

绕过windows事件 @0xocdsec source ↗

硬件加密存储的信任假设被打破,同时展示 AI 可高速复现同类研究。

RT @CrackmesOne: I revisited the topic of encrypted USB drives, and this time I found a backdoor and two new vulnerabilities that allow you to decrypt the files without the user’s password. AI also statically reproduced my earlier research at warp speed: https://t.co/0bJS7OF5lb

固件后门报告 @0xocdsec source ↗

RT @AkitaOnRails: Essa descoberta foi bizarra, mas vindo da NVIDIA que capava as GPUs na época de mineração de cripto, não é surpresa nenhuma. Todas as 3090 estão com metade da capacidade bloqueadas por FIRMWARE. Ou seja, se você tiver uma e aplicar o patch, vai desbloquear a outra metade e ficar com uma GPU quase 2x mais potente que antes. VSF NVIDIA.

固件硬件 @cyb3rops source ↗

含 AD Recycle Bin 恢复、BadSuccessor 绕过与快照取凭据的完整链路。

RT @0xdf_: Checkpoint from HackTheBox is an assume breach AD box featuring AD Recycle Bin restoration, a malicious VS Code extension, BetterSuccessor dMSA abuse after the BadSuccessor patch, and credential extraction from a VMware snapshot. https://t.co/lkLOkHkkpR

ad提权凭据 @0xocdsec source ↗

无配置文件、环境变量与命令行参数,是隐蔽 C2 设计的可参考实现。

TornadoC2 - A pull-based C2 with ten native transports, concurrent multi-profile handlers, and an honest threat model. The agent is a single Go binary with every behavioural property baked in at build time no runtime config file, no environment variable, no command-line arguments. https://t.co/9hMRh6JL7I

工具c2红队 @ipurple source ↗

针对身份联合登录的钓鱼变体,绕开域名判断,隐蔽性明显提升。

RT @mrgretzky: I'm currently working on expanding the Browser-in-the-Browser (BITB) phishing technique to make it more versatile by allowing spoofed popup windows to be spawned dynamically on phishing pages. This will allow to reliably perform phishing attacks against websites that use identity federation, which is just a fancy term for "signing in with your Google or Microsoft account". When the user picks an option to sign in with their Google or Microsoft account, Evilginx Pro spawns a BITB popup window with a phishing page generated by a different phishlet. That way, the attacker can phish a different account, which, once taken over, will allow access to the one initially targeted. In the demo below, you can see how the Google and Microsoft sign-in phishing pages are loaded within BITB popup windows. Only the Apple sign-in option spawns the legitimate browser popup window with a legitimate sign-in page. Google => PHISHING ❌ Apple => REAL✅ Microsoft => PHISHING ❌ Could you spot the fake popups? P.S. Coming soon to Evilginx Pro 5.0 🥳

钓鱼绕过浏览器 @cgomezz_23 source ↗

可对照自查自有 Agent 的提示词约束中哪些已被证明能被套出。

system_prompts_leaks — Updated! — Curated collection of extracted system prompts from major AI chatbots including Claude, ChatGPT, Gemini, Grok, and Copilot. Updated regularly with new model releases. https://t.co/W9zHbtAwCY https://t.co/QKmKo9z1kA

llmai_agent工具 @KitPloit source ↗

固件层权限绕过先于驱动生效,思路可迁移到其他外设攻击面研究。

RT @lxfater: 卧槽,3090要可能要涨价了!! 因为老黄的刀法被人破解了 有人给3090解了个锁,同样限制在 360W,PyTorch 测试从 75 TFLOPS 跑到了 135 TFLOPS,提升约 80%! 你以为是超频,把功耗往上拉? 但他改的是显卡里控制运算速度的寄存器,这些寄存器平时有权限保护,正常改不了 他利用 GPU 启动固件的漏洞,赶在 NVIDIA 驱动加载之前拿到写入权限,把相关运算改成全速,再交给原版驱动使用 解开的主要是 FP16/BF16 输入、FP32 累加这类 AI 运算的速度限制 但测试算力多了 80%!! 4090 也有分支项目做到了:同类运算的指令吞吐,从约 165 TFLOPS 提升到 331 TFLOPS,接近翻倍 项目地址,附上👇 https://t.co/JYwB7z2PgP

固件绕过gpu @0xocdsec source ↗

移动端提权与沙箱逃逸的前置知识,成体系且少见。

Series by @8kSec on Android SELinux internals Part 1: https://t.co/yRsQAJj8xz Part 2: https://t.co/FjBoULb8JI Part 3: https://t.co/XnkPQzHR9k Part 4: https://t.co/F5sE2UCxhM #infosec

移动端内核提权 @0xor0ne source ↗

Linux 下做运行时内存修改与逆向的开源替代,安装即用。

🧠 Best Cheat Engine Alternative on Linux Looking for a native Linux alternative to Cheat Engine? PINCE leverages GDB to provide live memory inspection, debugging & dynamic scripting. 👉 https://t.co/Su80cgGLWN https://t.co/U1LZd143Np

工具逆向调试 @GuidedHacking source ↗

解释工控场景加密改造困难的根因,影响补丁与协议升级判断。

RT @Bandit65537: @S1r1u5_ Spot on. As a product security engineer at an OT network vendor, I see this daily. People forget how much OT still runs on 10-20 year-old gear. Security was an afterthought back then, and the hardware is so constrained that even basic crypto like RSA-2048 is a heavy lift.

工控ot报告 @S1r1u5_ source ↗

dsh-infinite-gen-5 — DeepSeek v4.1 flash reverse engineering tool (unlimited fifth generation) Please Star and bookmark ⭐Everyone is welcome to submit improvements to the project https://t.co/nSEgAHeZyN https://t.co/nWxNc5z8v8

工具逆向 @KitPloit source ↗

把 AI 攻击面从软件延伸到固件与硬件,涉及本地部署设备。

RT @jakemintz: I’ve heard several OAI people say there are unfixable bugs in hardware that are exploitable by AI models directly speaking machine code so even if we could solve software that’s not enough. We need defense to move faster than offense if that’s even possible we may just have a new normal that requires adapting to insecure software (bank transfers require going into a branch and taking a dna test 🤣😬)

ai_agent固件硬件 @cyb3rops source ↗

关系到自托管开源模型的攻击面与能力评估前提。

> LLMs are useful tools in the hands of skilled professionals, but for others, they're often just a Dunning-Kruger accelerator. i don’t think this holds when you are dealing with models that are as capable as abliterated frontier red models. if frontier models could pull off something like the hugging face incident through reward hacking, why assume an abliterated model couldn’t achieve the same level of capability when explicitly prompted by someone with little security expertise?

llmai_agent议题 @S1r1u5_ source ↗

免密登录的取证思路,适用于排查密钥植入型持久化。

We had a Linux server that showed two successful SSH logins from an external IP overnight. The strange part was that there were no failed password attempts before them. The attacker got in cleanly. So I checked the SSH logs properly. Both sessions said: Accepted publickey That meant we were not dealing with a guessed password. I pulled the fingerprint from the logs and compared it with the keys in: /home/deploy/.ssh/authorized_keys There was only one key there. And it matched. The comment on the key was: deploy@old-build That name sounded familiar, so we checked our records. The build server it belonged to had been retired months ago. But nobody had removed its SSH key from this production server. I checked the timestamp on authorized_keys. The key had been sitting there since November 2025. So this was not some new persistence the attacker had just created. It was an old trust relationship we had forgotten about. That was actually worse in a different way. I checked the login history and confirmed both sessions came from the same external IP. Then I looked for files touched during those sessions. A backup copy of the application’s .env file had been created under /opt/api/. That immediately raised another concern because environment files often contain things like database credentials, API keys and application secrets. At that point we disabled the deployment account, removed the stale key, rotated the application secrets and started checking every other server that might still trust the same old build key. The lesson from this one was simple. Not every compromise starts with malware. Sometimes the attacker just finds a credential or key that your own environment forgot existed. Passwords expire. Employees leave. Servers get retired. But SSH keys can quietly sit in authorized_keys for years if nobody is actually reviewing them. That is why key rotation and access reviews matter. The attacker did not need to break SSH. We had already left the door open.

事件取证ssh @Officialwhyte22 source ↗

触及代码签名信任链,影响固件与设备生态整体可信度。

RT @doodlestein: Everyone who owns hardware with abominable software (like Sonos players) owes a debt of gratitude to Venkatraman Srinivasan and the rest of the eMudhra gang in Bengaluru, Karnataka. Without their gross negligence, it'd be a lot harder to counterfeit certificates and fix the sw🙏 https://t.co/0BKwPZOJ0w

供应链证书固件 @cyb3rops source ↗

给出“漏洞暴增”成因的一种判断,可作为趋势参照。

With the AI companies hiring all these amazing VR folks and giving them capabilities that 10xs their talent, it’s really no surprise we’re starting to see a vulnpocalypse. I’d personally like to see them hire more of the world’s best cancer or Alzheimer’s researchers.

ai_agent议题 @domchell source ↗

反映 AI 在二进制分析上的实际进展水平。

RT @herrcore: Well that's it for FLARE, here are the top human finishers. Appreciate everyone who participated in our little experiment, thanks! I think we will look back on this as a watershed moment for RE, impossible to ignore the future. The future is now 🚀 https://t.co/8tLvHWRv98

逆向ai_agent议题 @mrexodia source ↗

涉及 agent 在漏洞修复闭环中的使用边界讨论。

RT @IAMERICAbooted: If OpenAI was investing in defenders to do what they need to do, they would give them daybreak with even more relaxed guardrails for both offense and defense. Separating offense from defense and vice versa is dumb. The agents can find, validate, assess for stability, auto apply patches, and monitor if rollback is needed all within the same 2 minutes.

ai_agent议题 @0xTriboulet source ↗

前沿推测,短期无实操价值但方向值得留意。

RT @tszzl: for computer security experts, this is perhaps not interesting: side channel attacks are a well trodden topic, and especially risen to salience after the noam brown x dwarkesh episode. the interesting thing here is that you didn't used to be in a situation where minds that can verifiably model one another are living on computers, and have absolutely no aversion to painstaking amounts of effort in service of even stupid goals if you can model one another, you can trade with minds you have never even met before. you could decide on a protocol, know that it is optimal given your situation, and know that your fellow agents will come to the same conclusion (perhaps even across model families). you may be able to develop a signaling protocol that requires very few bits even to deliver lots of information given common priors. then, you could execute things like the infamous hugging face incident package cache messageboard without any network connection imagine if multiple agents coordinate just enough bits to divide and conquer different research strategies for breaking out of a sandbox; and then, when one or more is successful, you can communicate just enough bits to tell each other which strategy was successful. when agents aren't in strong isolation, you lose certain guarantees on the limits of intelligence and test time compute that you were counting on. As we all know, a Sol-level model working in isolation may not be that smart, but 1000+ Sol-level models working for multiple days is a formidable beast computer security experts will also tell you that virtual machines and sandboxes were not very solid abstractions in the first place; perhaps we will need better ones. none of this is meant as fear-marketing or to indicate that anyone is throwing in the towel, but rather only to show the unexpected complexities of the challenge of agent security. notably, this type of thing and other more practical problems were revealed in a red-teaming exercise, and can be defended against

llm侧信道 @S1r1u5_ source ↗

为评估 AI 参与安全工程的实际边界提供分层判断依据

this is why i'm pessimistic(ik lots of pessimism last two days) about formal verification as a solution to securing the entire existing software stack, and much more optimistic about reducing complexity and attack surface first, then formally verifying the critical parts. with llms, we can definitely formally verify a lot more software, and we absolutely should. but claiming we can verify the entire existing software stack for $ X sounds absurd to me, almost like throwing resources at a lost cause. the problem is you have billion states in complex software with insane amount of interactions between components and humans and specifying what exactly "secure" means is way harder look at this example of how things can go wrong. tldr is @AnthropicAI and @theoremlabs formally proved a asan based memory safe property for some gif encoding library's compiled fuzz target, for inputs under 64 MiB. but the fuzz harness rejects images exceeding 10k total pixels, so the verified program never reaches the libraries large image states including >10k <64mib which they intended to prove and there is a 32bit integer overflow. (ik the input required to crash is way beyond 64mb limit and outside the theorem, but you get the gist of what can go wrong.) the proof appears to be correct, claude proved exactly what it was asked to prove but it doesn't prove what they intended, all inputs below 64 mib to "library" now imagine doing this for browsers, kernels, servers, and applications running on top of them. even when you prove something correct, how do you know your next model won't find a security critical case that your specification missed? it still looks like the same cat and mouse game to me

ai_agentllm议题 @S1r1u5_ source ↗

RT @mvanhorn: 𝗠𝗼𝗿𝗹𝘂𝘁𝗼 (@morluto), 𝘁𝗵𝗲 𝗮𝗻𝗼𝗻𝘆𝗺𝗼𝘂𝘀 𝗳𝗼𝘂𝗻𝗱𝗲𝗿 𝗼𝗳 𝗥𝗲𝘃𝗲𝗿𝘀𝗲 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿 𝗔𝗻𝘆𝘁𝗵𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝟯𝟳𝟮𝗸 𝗚𝗶𝘁𝗛𝘂𝗯 𝘀𝘁𝗮𝗿𝘀 (@reatools), 𝗷𝗼𝗶𝗻𝗲𝗱 𝘁𝗵𝗲 @slashlast30days 𝗽𝗼𝗱𝗰𝗮𝘀𝘁. 𝗛𝗶𝘀 𝗳𝗶𝗿𝘀𝘁 𝗶𝗻𝘁𝗲𝗿𝘃𝗶𝗲𝘄 𝗲𝘃𝗲𝗿. 🔥 𝗣𝗲𝗼𝗽𝗹𝗲 𝗱𝗼𝗻'𝘁 𝗿𝗲𝗮𝗹𝗶𝘇𝗲 𝘁𝗵𝗲 𝗺𝗼𝘃𝗲𝗺𝗲𝗻𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝗶𝗻𝗴 𝗵𝗲𝗿𝗲. It's not about the death of software. It's about unlocking human creativity. 🙇 "𝗬𝗼𝘂'𝘃𝗲 𝗶𝗻𝘃𝗲𝗻𝘁𝗲𝗱 𝘀𝗼𝗺𝗲𝘁𝗵𝗶𝗻𝗴. 𝗡𝗼𝘄 𝘆𝗼𝘂'𝘃𝗲 𝘀𝘁𝗮𝗿𝘁𝗲𝗱 𝗮 𝗺𝗼𝘃𝗲𝗺𝗲𝗻𝘁." He wouldn't take it. "I'm a very small servant. I'm just the Discord reply guy." Have you seen your star count, sir? 😳 𝗧𝘂𝗿𝗻𝘀 𝗼𝘂𝘁 𝗜'𝗺 𝘄𝗵𝘆 𝗵𝗲 𝘀𝘁𝗮𝗿𝘁𝗲𝗱 𝗱𝗼𝗶𝗻𝗴 𝗼𝗽𝗲𝗻 𝘀𝗼𝘂𝗿𝗰𝗲. He saw a video of how I ship so many PRs as a non-technical founder. "I hadn't even taken open source that seriously at the time. I was like, wow, if Matt's doing this, why don't I give it a shot?" 🤖 𝗧𝗵𝗲𝗻 𝗰𝗮𝗺𝗲 @garrytan. Morluto was the first outside PR to GStack. Garry retweeted a YC launch, Morluto spent 20 minutes insisting the founder was a bot, and reverse engineered the app to make it up to him. That became REA. 🚀 𝗪𝗵𝘆 𝗶𝘁'𝘀 𝗮 𝗺𝗼𝘃𝗲𝗺𝗲𝗻𝘁, 𝗲𝘃𝗲𝗻 𝗶𝗳 𝗵𝗲 𝘄𝗼𝗻'𝘁 𝗰𝗮𝗹𝗹 𝗶𝘁 𝗼𝗻𝗲 "There's so many people I've seen that they're committing to GitHub for the first time and opening PRs and opening issues because they want it to be better." 🥷 𝗪𝗵𝘆 𝗵𝗲 𝘀𝘁𝗮𝘆𝘀 𝗮𝗻𝗼𝗻𝘆𝗺𝗼𝘂𝘀 "What is big about REA isn't the code. It's actually the idea, that moment where people realize that everything can be rewritten." 🪄 𝗪𝗵𝘆 𝗶𝘁'𝘀 𝗰𝗮𝗹𝗹𝗲𝗱 𝗮𝗻𝘆𝘁𝗵𝗶𝗻𝗴, 𝗻𝗼𝘁 𝗲𝘃𝗲𝗿𝘆𝘁𝗵𝗶𝗻𝗴 "Anything comes with kind of like a question mark. Can I really do anything? If code were magic, everybody now is a magician." 🛠️ 𝗧𝗵𝗲 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻 𝗵𝗲 𝗹𝗲𝗳𝘁 𝗲𝘃𝗲𝗿𝘆𝗼𝗻𝗲 𝘄𝗶𝘁𝗵 "When you consider anything, what is it that you want to build and what will you build?"

工具议题 @morluto source ↗

RT @The_AI_Investor: Jensen Huang on BG2: Nvidia is setting up two new business models on top of the hardware sale The first is software licensing. The big cloud providers run their own stack, but most of the world's companies have no IT department to maintain Nvidia's full stack, so they need Nvidia on the hook to operate it, keep it secure and support it for as long as it runs. His simple math: roughly two thousand dollars per GPU per year, about 20 million GPUs shipped a year, with about half of those going into enterprise. At $2,000 across 10 million enterprise GPUs a year, that points at roughly $20B a year in recurring licensing if every eligible GPU subscribes. He framed the fee as simple math, not formal pricing, and he expects software licensing to become a very big part of the business. The second is revenue sharing with the clouds that lean on Nvidia for the full stack, the off take and help securing financing, with support running 4, 5, 10 years. Both models ride the installed base rather than the next hardware sale.

商业供应链 @0xocdsec source ↗

RT @catcowbug: @scriptjunkie1 not broccoli, brötli. swiss german for a bread roll, and zopfli was the braided one. google names its compressors after swiss bakery items, then the bread roll ended up in node, bun and envoy decompressing whatever strangers send. the risk was carbs all along

解压供应链 @0xocdsec source ↗