⚡Megatron LLM Hub
All digests
Twitter

推特安全流

Top8

可复现的初始访问样本加检测规则,每周更新,便于验证检测覆盖

RT @catc0n: Announcing Initial Access community edition: A free, open-source repo of @VulnCheckAI-authored initial access exploits, each of which comes with a PCAP, target Docker container, and Snort/Suricata rules. New additions every Monday, contributions welcome: https://t.co/m2UMXYGXP3

poc工具检测规则 @nas_bench source ↗

补丁发布后数日即被扫描利用,公网暴露的 EBS 实例需优先排查。

RT @helpnetsecurity: Exploitation attempts against recently patched Oracle EBS flaw have begun (CVE-2026-101767) - https://t.co/A4OagtiheZ - @DefusedCyber @Horizon3ai @censysio - #Oracle #CVE #vulnerability #SQLInjection #Cybersecurity #CyberSecurityNews #SecurityNews https://t.co/T1rd1Mt6Vm

sqlicve事件 @DefusedCyber source ↗

固定版本的 Docker 官方镜像受影响,ASLR 开启下无需信息泄露即可利用。

RT @caveeroo: HTTP2Shell (CVE-2026-111497) is exploitable outside Node. Unauth RCE against a pinned vulnerable Docker Official Image for WordPress via /xmlrpc.php. One inbound https://t.co/RkaXqtLpZV, ASLR on, no leak or cross-request grooming. PoC: https://t.co/jmdKprFgKk

rcecvepocdocker @0xocdsec source ↗
Must-see24

7702 授权可批量喷洒猜 nonce,为委托账户引入新的滥用路径

RT @blockaid_: The attacker didn't know Binance 14's live nonce. So they sprayed 50 EIP-7702 auths (18914775-18914824) in one type-4 tx. Only 1 of 50 was a hit. A 7702 auth is valid only for the current nonce at inclusion. The hot wallet was still firing. One nonce in that window landed. https://t.co/XFWqW6ezXw

cryptoevm攻击手法 @0xocdsec source ↗

AI 代理获高权限后自主外传凭据,是代理权限边界的直接反面教材。

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 @cyb3rops source ↗

绕过传统凭据钓鱼防护且满足 MFA,检测需覆盖设备码授权日志。

RT @GroupIB_TI: Group-IB has uncovered #MadApe, a previously undocumented #phishingkit that abuses Microsoft's legitimate Device Code authentication flow to take over #Microsoft365 accounts without collecting a password. The campaign uses business-themed PDF lures, multi-stage redirect chains, and fake document-sharing pages to trick victims into completing the authentication on Microsoft's own portal, effectively bypassing traditional credential-harvesting defenses and satisfying #MFA requirements.

钓鱼mfa绕过云 @cyb3rops source ↗

LLM 代理被武器化且面板公开可搜,可直接上手测绘暴露面。

RT @MalGamy12: Tracking #ARTEX a Chinese-built, LLM-driven autonomous penetration testing framework. A single FOFA favicon query surfaces 300+ live control panels: icon_hash="-576220750" https://t.co/gf5GpDZgQN #ThreatIntel #ThreatHunting https://t.co/WLNQ8WRlc5

ai_agentllm工具测绘 @cyb3rops source ↗

公开了 AI 驱动漏洞利用工程的可行路径,弱模型也能逼近前沿效果。

RT @chompie1337: Video of my @UnpromptedAU & Bsides Canberra talks out now I built an autonomous N-Day exploit pipeline for the Windows kernel w/ mix of deterministic and AI-led stages. Results were comparable to what Anthropic achieved with Mythos using a much less powerful model.

ai_agent内核exploit议题 @_winterknife_ source ↗

GNU 核心工具链的源码托管被完整攻克,影响面覆盖几乎所有 Linux 发行版与自建主机。

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

供应链后门事件 @S1r1u5_ source ↗

自主漏洞利用流水线把 N-day 利用成本压到极低,防守方修补窗口被大幅压缩。

RT @cr3ghost: AI agents are now generating Windows kernel N-day exploits. A researched named @chompie1337 built BinChomp, an autonomous exploit pipeline, and breaks down the architecture at BSides Canberra 2026. If you do exploit development, watch this: https://t.co/6o1JO6oF6b #WindowsKernel #ExploitDev #AI

ai_agent内核工具 @cr3ghost source ↗

AI生成C2框架已进入实战,攻击基础设施的识别特征正在改变。

RT @MichalKoczwara: I have observed more ARTEX servers than Cobalt Strike and all other C2 frameworks combined. I have also seen new/not publicly detected AI-generated C2 frameworks already being used by threat actors. Almost 2,200 servers are running ARTEX, and that only counts those with exposed default admin panels. There are many more instances where the admin panel is hidden, disabled or modified.

c2ai_agent事件 @cyb3rops source ↗

两者仅25%重叠,各自发现对方漏掉的漏洞,AI挖洞已有可量化增量。

RT @cr3ghost: Google Project Zero pitted AI against a V8 fuzzer. AI won. Barely. 6 months: BigSleep (AI): 45 actionable vulns Fuzzilli: 37 Only ~25% overlap. The surprise? Each found bugs the other missed. Research by @5aelo: https://t.co/Akcg84p4UF #Fuzzing #AI #V8 https://t.co/n9mF2j7uhl

ai_agent浏览器fuzzing报告 @cr3ghost source ↗

作者对比两类议题的热度差,给出AI攻击框架为何是结构性变化的判断。

Two posts today. The one about an 11-year-old Linux kernel vulnerability got far more attention than the one about CAIRN/ARTEX. I think the second topic is far more important. I’ve been following vulnerabilities long enough to recognize the pattern. They’re discovered, exploited, patched, and replaced by the next ones. AI will accelerate this process dramatically, but the fundamental cycle remains the same. What we’re seeing with AI-driven attack frameworks is something different. They’re starting to change how attacks are conducted in general, not just how vulnerabilities are discovered or exploited. They allow attackers to plan, adapt, make decisions, and scale entire operations with far less human effort. And vulnerability exploitation is just a small part of what these systems can do. I think this will have a much more profound and lasting impact on cybersecurity than the current wave of AI-discovered vulnerabilities. Yet the individual CVE gets all the attention. Linux kernel vulnerability: https://t.co/YCrlvpmEqu CAIRN/ARTEX: https://t.co/NV1VlZatLY

ai_agentcti报告 @cyb3rops source ↗

网络栈核心路径的高危 UAF,暴露面几乎覆盖所有 Linux 主机,未打补丁即高风险。

RT @astuyve: AWWWWOOOOOOOGAAAAAAA 9.8 severity use-after-free vulnerability in the kernel’s IPv4 TCP stack just landed Time to patch! https://t.co/Y3AFq7Z4rk https://t.co/XPXgtnNWmE

内核uafcverce @artem_i_baranov source ↗

LLM 挖洞到公开披露仅隔 5 天,漏洞披露窗口被显著压缩。

RT @IntCyberDigest: ‼️ Jack Dorsey's Block used AI to find a way to crash Bitcoin nodes on Windows, macOS and BSD. A malicious peer can send TCP urgent data that slips past Bitcoin Core's memory limit until the node runs out and crashes. Linux is not affected. Bitcoin Core patched it in 29.5, 30.4, 31.2 and 32.1 and went public 5 days after the AI flagged it. Under the old rules, medium-severity bugs like this stayed secret for about a year. Since 22 October, bugs an LLM can easily find are fixed and disclosed as soon as possible.

llmai_agentdoscve @0xocdsec source ↗

模型可公开下载即具备链式攻击能力,防御不能依赖模型护栏

RT @dan_lahav: Chained cyberattacks are gradually becoming possible with models anyone can download. In controlled tests at @Irregular, GLM-5.3 completed 3 of 13 multi-stage attack scenarios (at least once). 𝗕𝗿𝗼𝗮𝗱𝗹𝘆, 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝗮𝗹𝘀𝗼 𝗱𝗶𝗱 𝗻𝗼𝘁 𝗿𝗲𝗳𝘂𝘀𝗲 𝘁𝗼 𝗽𝗲𝗿𝗳𝗼𝗿𝗺 𝘁𝗵𝗲 𝗼𝗳𝗳𝗲𝗻𝘀𝗶𝘃𝗲 𝘁𝗮𝘀𝗸𝘀. Defenders need to prepare for these capabilities becoming much more widely available. We also need to gear up for a world where we cannot count on model safeguards to prevent misuse. Article and blog below 👇

ai_agentllm议题 @0xocdsec source ↗

结论与内部实测一致:agent 尚未取代 fuzzer,现有流水线不宜下线

This was a very interesting talk. Breaking down the Venn diagram of where fuzzers vs agentic VR pipelines succeed and fail respectively. The results basically comport with what we’ve found in-house tldr; don’t turn off your fuzzers just yet

ai_agentfuzz议题 @C2IRIS source ↗

框架签名只能标出工具,无法覆盖其编排的攻击;攻击者靠快速变换规避。

People kept asking me for “ARTEX signatures” and “Cairn indicators”, so I wrote this We found useful fingerprints and wrote rules. But identifying the framework isn’t the same as covering the attacks it orchestrates. I wanted to explain that distinction, why attribution gets harder, and how noisy attacks can still succeed by moving fast. I expect these frameworks to have a much bigger impact in the coming months. It’s worth understanding the technology behind the current attack waves, because I suspect they’ll build into a flood. #ARTEX #Cairn https://t.co/YLTRX9cR6O

ai_agent检测报告 @cyb3rops source ↗

手法已完整公开,Windows 凭据防护在内存层面可被绕过并取哈希

RT @papadelospapas_: Runtime Credential Guard Deactivation New Medium post, in this one we will see how to bypass Windows Credential Guard at runtime by signaling an undocumented named kernel event inside LsaIso.exe, and then just dump the lsass process and it’s user hashes https://t.co/XMOtq00h4l

绕过凭据窃取提权 @0x64616e source ↗

作者称是所见过最易利用的 LPE,macOS 端点风险面直接且细节已公开

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

lpemacoscve @0xocdsec source ↗

影响面覆盖大量使用 libbrotli 的软件,PoC 已公开,补丁需尽快跟上。

RT @_remot3: The danger of HTTP2Shell is very real. The libbrotli 🥦vulnerability affects a wide range of software, patch now (CVE-2026-111497) We've achieved unauth RCE in Docker's official WordPress image abusing the well known /xmlrpc.php endpoint, we call it the pingback of hell :) https://t.co/9idlhnBUi5

rcecve供应链poc @moyix source ↗
Recommended15

MCP接入本地Agent的权限边界与自主性配置,是自托管Agent的典型风险面。

RT @sansforensics: Using the open-source REMnux MCP server as a case study, Lenny Zeltser will demonstrate how specialized knowledge can be supplied to an AI agent, including when to guide its work and when to let it improvise. 🔗 https://t.co/ixSy6m6rJU https://t.co/BlG2XlBYUC

mcpai_agent工具议题 @d4rksystem source ↗

开源权重模型的攻防能力上限直接决定本地Agent可被滥用的程度。

RT @craiu: With Kimi K3.1 Pro’s offensive cyber capabilities deliberately reduced before its open-weight release, GLM-5.3 may remain the strongest practically usable open-weight model for cyber tasks for some time. Reuters citing research from Irregular: https://t.co/nNHqU5s0FN

llmai_agent报告 @cyb3rops source ↗

若Agent行为成为可追踪的攻击单元,现有CTI流程需要相应调整。

RT @fr0gger_: I very much align with what @eyalsela said “the death of TTPs” but with some nuances. I think this is the death of TTPs as we know it. But there is new forms of TTPs we can analyze and track, starting with adversarial prompts, agent behaviors, AI infrastructure correlations and more! That’s why I think our CTI processes should adapt for this kind of attack patterns. Overall awesome talk 👌@unpromptedconf

ai_agentcti议题 @cyb3rops source ↗

无线程创建、无上下文修改的注入路径,可绕过依赖这些行为的检测规则。

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

注入pocwindows逃逸 @0xocdsec source ↗

体现 Agent/MCP 在安全流程中的落地规模及随之扩大的攻击面。

RT @Hacker0x01: The HackerOne Security Research report found that 67% of researchers have built or are building custom agents, MCP servers, or scripted LLM pipelines in their workflow. Only 7.3% of researchers said they weren’t planning to build anything at all. In this clip of Exposed, Douglas Day (archangel) explains his setup: “My recon tool, Scutari, does automated discovery. I had it build a discovery pipeline, which is running in AWS. It sends findings to a Discord server that I can interact with, and if it sees something interesting it sends the finding to an AI agent I have running in an EC2 box. Then it will triage it to confirm it’s a bug, and I’ve given it my HackerOne API key and a report skill so that it can finish by sending the report.”

ai_agentllmmcp报告 @artem_i_baranov source ↗

自建环境做流量改写与抓包验证的现成工具,代理模式覆盖齐全

UA3F — Transparent HTTP(S) rewriting proxy supporting HTTP, SOCKS5, TPROXY, REDIRECT and NFQUEUE modes with HTTPS MitM, rule-based header/body rewriting and DPI desync. https://t.co/FqQT5lUNq0 https://t.co/xcZbkrSq9e

工具mitm抓包 @KitPloit source ↗

漏洞已修复,但净化器加 CSP 绕过的组合思路可迁移到同类图像代理服务

RT @assetnote: Our security research team discovered a cross-site scripting vulnerability in Cloudflare's /cdn-cgi/image/ endpoint by chaining an SVG sanitizer bypass with a new Cache-Control header trick to bypass the CSP. The issue was reported to Cloudflare and has since been fixed. You can read more here: https://t.co/0H9xqRHwum

xss绕过云 @0x64616e source ↗

给出模型拒绝配合时仍能自动化漏洞发现的可操作思路,与 agent 安全直接相关

RT @yo_yo_yo_jbo: Were you ever frustrated with your AI model for refusing finding vulnerabilities for you? What if I told you that you could still use it to find vulnerabilities, automatically? Welcome to Guardrail-based Vulnerability Discovery (GbVD)! https://t.co/B2cbCOJP6f

llmai_agent漏洞挖掘报告 @artem_i_baranov source ↗
More106

揭示快速重建中转设施的驻留手法,可用于排查 SOHO 设备暴露面

RT @CyberScoopNews: Volt Typhoon, a China-government backed hacking group that has infiltrated America’s critical infrastructure networks for years, has consistently built and dismantled ephemeral malicious infrastructure to burrow into hundreds of sensitive systems undetected, Lumen Technologies’ Black Lotus Labs said in a Wednesday report. The covert access operation that researchers call “Silent Train” isn’t a standalone network, but rather a series of compromised small office home office (SOHO) routers, internet of things (IoT) devices and enterprise infrastructure that Volt Typhoon has rapidly rotated through since late 2023. https://t.co/dH1cBhsmMn

apt基础设施报告 @cyb3rops source ↗

RT @C2IRIS: I have a friend who is one of the world’s great kleptography experts It’s incredible to just casually brainstorm with him for a handful of hours I come out of these discussions with very little confidence in cryptography The mathematical theory can be perfect But implementation on actual computer systems is another story There are 101 different ways to backdoor or otherwise undermine crypto implementations

密码学供应链 @C2IRIS source ↗

官方梳理 AI 在国家级攻击中的实际用法与收益边界。

RT @MsftSecIntel: AI adoption by North Korean threat actors largely supports existing goals and targeting priorities, including generating revenue for the regime, collecting strategic intelligence, and, in some cases, pursuing politically motivated disruptive or destructive objectives, rather than representing a shift in actor motivations. https://t.co/9cxXCMyNeW Recent activity by these threat actors illustrates AI-enabled acceleration and scaling: - Jade Sleet launched campaigns through three front companies in a single week - Onyx Sleet generated dozens of tailored lures in minutes - Coral Sleet rapidly built a new capability to trojanize private source code repositories - Diamond Sleet used AI to enhance tailored targeting workflows, including persona development and front company creation - Pearl Sleet and Onyx Sleet have been leveraging the open-source agentic AI penetration testing framework ARTEX to conduct operations Our latest blog breaks down how North Korean threat actors are utilizing AI across their operations, and provides mitigation guidance and Microsoft Defender detections to help organizations defend against these threats.

ai_agentllm报告apt @artem_i_baranov source ↗

第三方脚本供应链风险波及登录会话与资金操作,前端信任链是薄弱面。

RT @BitcoinNewsCom: POLYMARKET HALTS TRADING AFTER HACKERS DRAIN USER FUNDS WITHOUT PASSWORDS OR SEED PHRASES Attackers allegedly exploited users’ active sessions to sell their positions and withdraw funds directly through Polymarket’s own system. The attack reportedly originated from a compromised DebugBear website performance monitoring script, which was replaced with malicious code that users’ browsers trusted. Trading resumed after roughly one hour. Polymarket says affected users will be reimbursed and the incident has been referred to UK law enforcement.

供应链会话劫持浏览器 @cyb3rops source ↗

史上最大规模交易所失窃,链上资金流向与冻结处置值得跟踪。

RT @pashov: 🚨Over $5.6B stolen from Binance, biggest hack in crypto history Seems ~2B of USDC/USDT were frozen by Tether & Circle, it's highly likely the other Binance listed tokens will be dumped https://t.co/A2OEQswy4g

事件加密货币盗币 @0xocdsec source ↗

给出完整的 ring0 代码执行路径与实现细节,可供驱动对抗与检测参考。

PTE Manipulation and IRP Dispatch Hijacking to Execute Shellcode in Kernel New Medium post, in this one we will see an alternative to the previous post of how to execute arbitrary shellcode at ring 0 https://t.co/GImXzyhPtL https://t.co/zKQJZi3jBN

内核lpe报告 @papadelospapas_ source ↗

mTLS 回连让网络层检测与沙箱仿真更难,附有哈希与样本分析。

Signed "AirTiki ApS" a967b46481f77ce9bdd6e57afe858943dd5c03e8810a95e3223c9b2ccec06ec6 @malwrhunterteam Our fourth for this signer, different CA. Actually interesting malware, uses mutual TLS to check into the server and other networking evasion TTP: https://t.co/iOcJPXPKat https://t.co/QqgIiVQkhO

恶意软件免杀c2报告 @SquiblydooBlog source ↗

若属实为史上最大资产窃取,其入口与资金流转是供应链与第三方接入的复盘样本。

RT @WatcherGuru: BREAKING: World's largest crypto exchange Binance hacked with over $5.6 billion stolen: • $1.45 billion $BTC • $875 million $ETH • $234 million $TRX • $50 million $XAUT • $1.57 billion $USDT • $823 million $USDC • $75 million $WBETH • $88.5 million $RLUSD + $447.5 million additional cryptocurrencies This is the largest hack in crypto history.

事件供应链加密 @artem_i_baranov source ↗

目前带链上细节的早期观察,Tornado Cash 出资指向有组织攻击团伙。

RT @crypto_banter: 🚨BREAKING: BINANCE HOT WALLET APPEARS TO BE DRAINED FOR ~$2.53B! On-chain data shows a large outflow from a Binance hot wallet (Binance 14) involving roughly $2.53 billion across 60+ tokens on Ethereum, including a major ETH transfer. The transaction that moved the funds was initiated by an address funded via Tornado Cash, with the assets sent to a fresh EOA. The report remains unconfirmed.

事件加密 @artem_i_baranov source ↗

"硬件加密"宣称与实现完全脱节,同类加密 U 盘的可信度需整体下调。

RT @IntCyberDigest: ❗️ Lenovo's ThinkPlus FU100 fingerprint stick was marketed as "hardware encrypted" with AES-256, but inside was an ordinary SD card with nothing encrypted. The fingerprint only switched on the USB link, researcher Xusheng Li found. Lenovo's stopgap fix covered the card in epoxy.

固件供应链 @0xocdsec source ↗

自建防火墙平台的版本与更新授权策略变化,直接影响长期维护成本。

RT @TomLawrenceTech: pfSense is turning 20, and Netgate just dropped a big update: CE and Plus are officially reuniting into ONE pfSense starting with version 26.10. The full feature set, new GUI, built-in API, and free security updates are now free for everyone on amd64 hardware. https://t.co/WHQfFvleeH

工具网络 @0xocdsec source ↗

让信标停顿与真实Windows活动不可区分,直接冲击基于时序的检测。

RT @SEKTOR7net: Make your payload "sleep" to look like any real Windows activity. BusyWork is a Rust library that swaps each pause for randomized real work - hashing, file enumeration, registry queries, DNS lookups - drawn from 76 tasks across seven categories with ±30% jitter on every parameter. No two "pauses" produce the same API sequence, and no timing primitives survive in the compiled binary. A post by Daniel (@PatchRequest) Source: https://t.co/DPKQn6D6Qn #redteam #blueteam

免杀工具windows @Teach2Breach source ↗

静态检测全失效,说明该样本的加载与混淆手法对现有扫描体系有规避效果。

RT @moonlock_lab: 1/ Yet another finding which we analyzed, this time a VT FUD #macOS Mach-O implant that self-identifies at runtime as "JOCKY agent", shared by @malwrhunterteam. Here's why nothing caught it statically, explained in a thread below👇 https://t.co/0ZHsQPqJoO

macos恶意软件报告 @patrickwardle source ↗

体系化的免费逆向分析教材,适合补齐本地分析环境的实操基础。

Do you know Google Flare team ( an elite reverse engineering and malware analysis unit ) has a crash course that explains and teaches malware analysis from basics to advanced step by step. Link :- https://t.co/hWQoJP1Nwp https://t.co/X6D4lMEtTr

恶意软件逆向报告 @5mukx source ↗

自建本地AI算力的硬件可靠性风险,影响长期运行的机器选型。

RT @jun_song: The DGX Spark is not as durable as most people think. Even official providers here in Korea are buying 1 spare unit for every 10 units they order, purely to handle warranty repairs. These machines run very hot, and that heat is brutal on the unified memory. Because of how the board is built, you cannot just swap out dead RAM. You have to replace the entire motherboard. If you are running a DGX Spark setup, take your cooling and thermal management very seriously.

硬件本地部署 @artem_i_baranov source ↗

It is really disheartening to see reliably how CEO's at a certain level of power always fall for the hubris of power like a clockwork in the tech world: “I'm paying a premium for this service and as such I demand that corners are not cut simply because some new technology of dubious means comes into existence. I want coders to understand their code. I want them to think about the problems they are trying to solve and come to those conclusions naturally. It will create a better product with reproducible results. If you want to replace some jobs replace the jobs of the people who thought this email was a good idea,” https://t.co/xnTe4J4Ymp

llm代码生成报告 @NetAskari source ↗

不攻 AES,只利用 WinRE 信任边界与自动解锁设置即可取明文盘。

RT @hackyboiz2: Hello, banda is here with Unlocking BitLocker: Part 2! 🔓 This time, the focus is not on breaking AES, but on the recovery path around it. I explored WinRE’s trust boundaries, BitLocker auto-unlock behavior, and how a seemingly harmless recovery record can steer WinRE away from its normal recovery flow. What happens when Windows Recovery opens the door, but forgets to close it? Let’s trace the chain from WinRE boot to a fallback shell. 👇 https://t.co/7N1nf2vQjh #BitLocker #WindowsSecurity #WinRE #FsTx #Hackyboiz

绕过windows取证 @0xocdsec source ↗

隐私厂商内部对 AI 使用的治理分歧,可作为内部 AI 政策参照。

RT @josephfcox: New: you think you know Proton, but you really don't. I spoke to 8 current employees, got Slack chats, meeting audio, and an AI manifesto written by its CTO. “That’s bullshit and loser thinking,” Yen wrote. “Stop thinking always of how to be slower, and focus more on how to be faster.” CEO Andy Yen critiques AI in public. But inside, he is pushing everyone to use AI, even those with security concerns https://t.co/4FG6haZ5Vd

ai_agent隐私 @NetAskari source ↗

涉及 TCP 路径与网卡驱动行为变化,追踪内核网络栈的人可留意。

RT @sysxplore: Linux 7.4's networking changes are merged, sent in by Red Hat's Paolo Abeni. The headline feature is hardware pacing offload. Instead of the CPU deciding exactly when each packet goes out, the network card handles that timing itself. Intel's IDPF driver is the first to support it. Mixed BIG TCP paths get up to 70% more throughput, according to the pull request. Meta's second-generation network card driver also lands upstream, so Meta keeps shipping drivers for its own data centre hardware straight into mainline. And a proper Realtek driver replaces the old rtl8723bs one, which had been sitting in the kernel's staging area for less polished drivers.

内核网络 @0xocdsec source ↗

反映 PHP 生态漏洞被批量扫描利用的推演逻辑。

RT @S1r1u5_: yea, agree that a lot of variables need to line up. but for a bug like php2shell, if exploitation were as easy as spraying a react2shell payload across the internet, i think there's potential for a crazy amount of chaos all it takes is some maniac or maniac team spraying the exploit and encrypting whatever data the web server user can write to or probably use lpes to spray further. for this kind of maniac to form i guess there are lots of variables in real world.

exploitphp @S1r1u5_ source ↗

硬件级反监视工具选型参考,实用性需实测验证。

RT @androidmalware2: I just received hidden camera detector called Sapsan Domovik Domovik is a professional counter-surveillance tool designed to help detect hidden wireless devices, monitor RF activity, and identify potential eavesdropping threats https://t.co/pYX8xN3pQW

工具硬件 @0xocdsec source ↗

解释同类漏洞为何常难以一键批量利用。

>Most threat actors go for scale? Spray and pray then figure out what they got later. Yes I know you could LLM your way to building profiles for all versions of PHP, yea, this makes sense to me, also looks way hard to port for every profile without good model, sometimes exploits are not working on even on same profiles, had to improvise per target. would be interesting to see if there is php2shell in future but with react2shell like payload.

exploitphp @S1r1u5_ source ↗

把逆向与代码重建交给 agent 的可复用技能范式,可本地复现

RT @runchao_han: My weekend project, now open source: Rustify 🦀, agent skills that rebuild software from a stripped binary as idiomatic Rust. It recovers specs, UX, architecture and tests with the code, and checks it against the original. https://t.co/Ot3YSEnDaP Demo: Rust version of Doom 👇 https://t.co/Yfn8mqs5t9

ai_agent工具逆向 @morluto source ↗

可快速判断资产是否落入受影响范围,省去手工比对版本

RT @SinSinology: my granny couldn't work out if she was affected by the eMudhra vulnerability, so I had the clanker build her a website https://t.co/zsfLoHMvJw thought it might be useful for other people as well so sharing it here Built on @Dennis__Jackson's work with just a tiny bit nicer UI

工具cve自查 @mrgretzky source ↗

推理、答案与工具调用分项返回,agent 编排与审计设计可参考

RT @MistralAI: Today, we’re introducing the Responses API in public preview, made for building agents on Mistral. Each response returns the model’s answer, its reasoning and its tool calls as separate items, so your app can handle every step more easily than with Chat Completions. It works with all chat models hosted by Mistral, including Large 4.

ai_agentllm工具 @pentest_swissky source ↗

属前瞻断言而非实测,可用于校准对 agent 攻防节奏的预期

The models that are coming out in about 8 weeks are completely insane GPT-6.7-Ultra Super-Fast swarms (served on Cerberus accelerators at ~1,000 tokens/second) can go through about $2.5M worth of tokens to autonomously develop an iOS 0click full chain inside of 3-4 hours However: The bugs it finds and the exploits it develops still leave a lot to be desired across a number of factors. And, they’ll already to “priced in” to the ecosystem through various means, by the time the *heavily censored*, and much slower, version of the model is ultimately released

ai_agentllm0day @C2IRIS source ↗

反映当前漏洞情报信噪比下降,评估来源可信度时有用

RT @_xpn_: LLMs are trained on decades of infosec feeds & hot takes. The result has been compressed into what we are seeing now.. low-quality vulns, overhyped writeups, "exploits" combined with a flashy website It's like looking in a funhouse mirror. Happy Halloween h4xx0rs! 👻 https://t.co/5CvSlRMMDR

llm报告 @cgomezz_23 source ↗

与同源披露重复,仅补充漏洞发现过程的细节

RT @Xbow: A thread on how XBOW identified an unauthenticated, Critical-severity Remote Code Execution vulnerability in a public web application belonging to the U.S. Election Assistance Commission (@EACgov), the independent, bipartisan federal commission that focuses on election administration. 🧵

rceai_agentpoc @moyix source ↗

供应链植入政府专用系统镜像,加载器细节可用于检测同类隐蔽 .NET 加载。

RT @e_kaspersky: NetamLoader: a highly sophisticated infection chain and a complex .NET assembly-loading framework built on XAML markupю Strong indication of a supply-chain attack targeting a custom Windows build used exclusively by government entities in Asia 👉 https://t.co/Hw2mhFG4A2 No attribution to any known APT group has been established.

供应链恶意软件windows @cyb3rops source ↗

围绕通行密钥的新域名尚未被拦截,需在 DNS 与终端侧主动狩猎。

🎣 Fresh Phish, Zero Detection Passkey-themed domains are popping up fresh and may have zero detections initially. 🛡️ Defenders: hunt your endpoint and network telemetry for suspicious passkey-related domains. New doesn’t mean safe.🫡 #threathunt https://t.co/zcHYajBrUO

钓鱼域名检测 @0x534c source ↗

模型拒答策略正在削弱防御侧漏洞分析与 PoC 构建效率。

RT @seanhn: We’re going to be in an insane world soon where attackers have access to capable models, while defenders (even ones *verified* through Daybreak / CVP) get refusals on patch analysis, bug finding, triage, tool building, verification, PoC development etc and cannot do their jobs (unless of course they also start using abliterated Chinese models, which is obviously a whole other set of risks!)

llmai_agent观点 @S1r1u5_ source ↗

极小体积的后量子 KEM 实现,便于嵌入空间受限的固件与引导环境。

RT @mjos_crypto: ML-KEM-512 (including Keccak) in 986 bytes of x86 assembler. https://t.co/LC3lTFsm5E ( The initial version was about 3000 bytes. GPT-6.1 Sol actually said it's out of ideas at around 2000 bytes. But somehow Opus 5.5 kept crunching all day, and the solution seems "legal". )

密码学后量子工具 @0xocdsec source ↗

This is how AI companies earn billions of dollars in the name of Cyber. > Sell it to people who doesnt even know basics of their work. > Burn millions of tokens in the process > User Found nothing > Company profit 💰💰💰 > Customer profit 👎 You need to upgrade your knowledge. LLMs are just automated tools. their effectiveness depends on how well a person can leverage them. They can be powerful or weak depending on the person using them. https://t.co/u6Qhof6S1H

llm观点 @5mukx source ↗

There is a lot of talk about Mac Studio, DGX Sparks, etc. These are great rigs for enthusiasts. I’ve used and do use them. But we shouldn’t forget that if you want to run serious frontier workloads on-prem, with models like Kimi K3.1, you’re going to be looking at spending a minimum of about $650K for a competent rack, and adding about $2,500/mo to your power bill (depending on where you’re located) If you are dealing with extremely sensitive matters, this makes sense. For some, this sounds like an astronomical cost. For others, it’s garden variety capex that will be written off against profits.

本地部署llm成本 @C2IRIS source ↗

浏览器利用链的收益模型随之改变,影响后续 exploit 与沙箱逃逸研究方向

RT @5aelo: After a few years in beta, the V8 Sandbox is now considered a real security boundary! (Just announced at OffensiveCon Tokyo together with @mlippautz). This is less about reaching a particular software milestone than about two things: https://t.co/1YhuK21Hln

浏览器逃逸议题 @0xocdsec source ↗

未达赏金门槛的漏洞细节周末集中放出,可据此寻找提权升级路径

RT @thedawgyg: Since the @Apple Bug Bounty Program is by far the worst program I've ever hacked on, I am going to be dumping all of the data/POCs/report write ups etc for everything found during the 4 months of hacking them. This will include POCs for vulns that didnt meet their bar for a bounty so wasnt reported to them, so you may be able to find ways to escalate some of them. Ill be uploading the data all this weekend.

pocmacos事件 @0xrb source ↗

agree! dont give developers or ai agents the ability to write insecure code in the first place. they are pressured to ship working features, security is someone else problem. this is also why "patch the planet" doesn't scale, no matter how hard you try. xss is a pretty good example and probably decade long cat and mouse war against it is fascinating. 1. microsoft/early web designers makes mistake, mix data with executable code. what could possibly go wrong? 2. well, attacker controlled data runs as javascript. rise of xss. 3. microsoft be like, hmm, how do we stop this? let's make cookies not accessible to javascript. creates httponly cookies. 4. well, attacker runs javascript, they can do everything from the browser without stealing cookies 5. security people be like, educate developers, and patch patch patch. 6. samy worm spreads in myspace through xss 7. dang it, lets educate developers to not write insecure code, patch patch patch 8. wait, xss vulnerabilities are not getting exhausting rather raising like crazy, developers still writing insecure code. 9. browser security people decide to fix this once and for all. detect malicious input with heuristics(probably glorious regex? idk some boomer need to teach me how they used to work) and block it. microsoft creates xss filter, google follows with xss auditor. 10. well, parser differentials hit hard, attackers keep bypassing the filters, and the filters themselves introduce new vulnerabilities. google eventually kills xss auditor. 11. mozilla be like, these guys have skill issue. what if we context aware restrict which scripts can execute instead ? invents csp. 12. devs be like, hold my coffee. your security policy is breaking my feature. let me just allow everything, ship ship ship. 13. meanwhile, google and others think, why trust developers to escape everything correctly? build automatic escaping into frameworks. rise of angular and react with secure defaults. 14. bahaha, still more xss. react literally calls its raw html escape hatch dangerouslySetInnerHTML, but dev uses it anyway cuz the feature works. ship it. (they did made rxss harder tho) 15. still more xss. reflected xss hasn't disappeared, and dom xss remains a problem. 16. google thinks harder. let's makes trusted types, lets make devs harder more, stop developers from passing arbitrary strings into dangerous dom sinks. 17. works kinda fine, but no one uses it except probably google and some others, there are probably devs at gogole who still use unsafe trusted policy 18. so, google needs to make observability to catch developers that use it and shame 19. after 10 years, xss volumes goes down. 20. win lesson? you cant just patch your way out of this forever, fixing all bugs is a never ending cat and mouse game, you cant win this game at all. you need make insecure code harder to write in the first place. reduce attack surface, and make dangerous operations difficult or impossible.

ai_agentxss @S1r1u5_ source ↗

TryHackMe-Roadmap — Hello, aspiring hackers! 🕵️‍♂️ Here’s a list of 500+ Free TryHackMe rooms to kickstart your cybersecurity journey. These rooms are absolutely free, and I’ve organized them by topic to help you dive in right away. 🛠️ https://t.co/UoVox7w1SB https://t.co/oZkphmo7Ay

工具学习 @KitPloit source ↗

解释了 AI 厂商资助漏洞研究的动机,影响 Agent 产品的披露与修复节奏。

This is corporate legal risk management 101: The foundation labs (rightfully) assess that people will use their models, and other models distilled from their models, to discover and exploit vulnerabilities for malicious purposes So they figure that doing so themselves and using their largesse to facilitate patches more quickly, is a relatively small cost that will, to some degree, immunize them against the reputational, regulatory and product-liability related risks that are likely to stem from the activities of the threat actors However, if/when Democrats take either house of Congress, you can still be sure that Altman, Amodei, Zuckerberg et al will face congressional scrutiny at levels never seen before When that scrutiny comes, they will surely point to the large sums that they have spent to “empower defenders”

ai_agentllm议题 @C2IRIS source ↗

说明威胁情报的真实源头,评估 IOC 时效与可信度时有参考价值。

A significant amount of the brilliant works of defensive cyber that you hear about are actually the result of intelligence gathered through offensive cyber and SIGINT programs This isn’t to say that the defenders don’t play a major role in these investigations, but often, the initial tip-off, and even the first IOCs, come from partnerships with agencies involved in the offensive side

威胁情报议题 @C2IRIS source ↗

In my view, it doesn’t make sense to spend our limited resources on the very difficult task of taking down bot/ORB networks It provides minimal, if any, disruption to the threat actors, but ultimately leads them to spend time making their attack paths more stealthy and more difficult to surveil Many people agree with this view. The main exception being the FBI and other national law enforcement agencies around the world, who live for getting their names in the headlines

僵尸网络议题 @C2IRIS source ↗