必看18
家用路由器与 AP 固件更新滞后、难以加固,攻击面长期被低估。
Not to give you any ideas, but why has none of you looked for bugs in wifi firmware yet? 🤔 https://t.co/NdZ5mYzw3r
给出用LLM编排攻击流程的具体框架,可迁移到本地Agent红队
Automating AI Hacking with Tree of Attack - John Stigerwalt @WKL_cyber
https://t.co/2f3ySftb34
ai_agentllm自动化报告
@pentest_swissky
原文 ↗
说明Agent框架高度同质化,选型应看方法论而非品牌
People focus on the names, on the frameworks … forget the names. They’re all smoke and mirrors.
Anyone could task a coding agent with studying one of these frameworks and building a custom version in another language, with a different flavor, skill set or type of orchestration.
We’re in a new era now. Implementations are becoming disposable. Thousands of lines of code can be spewed out with very little effort. The principles and practices matter more than the framework names, because those ideas can materialize in code after a single prompt.
ai_agentllm议题
@cyb3rops
原文 ↗
iOS 最新两个大版本同时受影响,利用代码已放出,修复窗口为零。
RT @forcequitOS: Introducing worse_query, a zero-day sandbox escape for iOS 26 and 27, again.
In collaboration with @0xjohnny!
https://t.co/dOZF8oaaw0
逃逸移动端0daypoc
@0xocdsec
原文 ↗
权重公开可本地部署,进攻性 agent 能力已逼近受限商用模型。
RT @IntCyberDigest: ‼️ DeepSeek's free, downloadable AI model is almost as good at hacking as the model OpenAI keeps locked behind a restricted cyber programme.
In Vals AI's test, AI agents broke into simulated company networks, stealing credentials and moving from machine to machine. DeepSeek V4.1 Flash hit 73% of the objectives, OpenAI's GPT-6 Sol 81%.
A DeepSeek run cost under $3. A Sol run cost about $22.
On one of the two networks, DeepSeek came out ahead.
Its weights are public, so anyone can run it with no safeguards.
ai_agentllm红队基准测试
@artem_i_baranov
原文 ↗
带时间切分,可直接用于漏洞检索与利用类型分类评测。
hackersignal — 7.45M-document cyber threat intelligence corpus and benchmark linking hacker-forum, exploit and advisory text to 340K NVD CVEs, with temporal splits for retrieval and exploit-type classification. https://t.co/mBB2y1EimH https://t.co/vtFBfoeVYz
llm数据集工具cve
@KitPloit
原文 ↗
把 AI 攻击框架的控制台基础设施与威胁组织挂钩,是 AI 驱动攻防少见的可操作归因线索。
Correlated 60 days of network IOCs against ThreatBook intelligence feeds to identify Artex/CyberstrikeAI console infrastructure. The enrichment enables attribution insights, highlighting threat groups potentially associated with Artex/CyberStrikeAI-enabled offensive activity.
#Threathunting #Artex #AIOffensiveFramework
ai_agent威胁情报事件
@0x534c
原文 ↗
针对 Chrome App-Bound Encryption 的新绕过手法已成武器化工具,凭证窃取防御需同步更新。
Abe RAT signed "Kimberley Hankinson"
h/t @malwrhunterteam
"Abe" is likely short for "App-Bound Encryption". Uses hidden browser to decrypt passwords
Leveraged @malbeacon 's DeceptionPro to watch hands on activity.
Full blog: https://t.co/JEUIixYOUk https://t.co/dKlpDb87Hh
绕过窃密恶意软件报告
@SquiblydooBlog
原文 ↗
默认配置的自建 SSH 服务可能被无凭证登入,补丁已发需尽快核对版本。
RT @TheHackersNews: ‼️WARNING - OpenSSH issues emergency fix for a critical SSH authentication bypass on Alpine Linux.
A password longer than 256 bytes can unlock key-only accounts on affected servers, without knowing the actual credentials.
Here's how it works ↓ https://t.co/hOKc4hirLg
绕过sshcve
@artem_i_baranov
原文 ↗
攻击型 AI Agent 已进入实战并被滥用,暴露服务器给出可复用的失陷指标。
RT @Dinosn: Inside ARTEX: AI Hacking Tools, Exposed Servers and South Korean Bank Breaches https://t.co/CLt42CeMKs
ai_agentllm事件
@artem_i_baranov
原文 ↗
自托管 AI Agent 的攻击面可直接对照这些开源框架做威胁建模。
RT @ShanHolo: Beyond #Artex, what other #offensive #AI frameworks should the cybersecurity community be watching?
A few on my radar:
🔹 CyberStrike
https://t.co/gC5QSBF1gY
🔹 Cairn
https://t.co/t5C21RHN6U
🔹 PentAGI
https://t.co/ZzJHCY1qid
Have you tried any of these? Let me know!
ai_agentllm工具
@artem_i_baranov
原文 ↗
BGP 会话签名保护失效,补丁刚上 netdev,尚无 CVE 分配。
RT @sysxplore: France's cybersecurity agency just posted a fix for a hole in Linux TCP authentication, and every 32-bit kernel built with MD5 signing is exposed.
TCP-MD5 (RFC 2385) is the old signature scheme that protects BGP sessions between routers. Its successor is TCP-AO.
Jérémy Jean of ANSSI posted the v2 fix to netdev on Saturday. With TCP_MD5SIG=y and TCP_AO=n, a packet carrying a junk AO option hits a stub that returns success, so a socket requiring MD5 accepts unauthenticated data. The reverse config has the same bug.
Eric Dumazet's review was blunt: an attacker able to hit the window of an MD5-protected session "can therefore inject RST or data with a junk AO option, which is exactly what RFC 2385 is supposed to prevent."
TCP-AO depends on 64BIT, so "every 32-bit kernel with TCP_MD5SIG=y is in the vulnerable configuration." The fix is tagged for stable, reviewed by Dumazet and Kuniyuki Iwashima.
认证绕过内核网络协议cve
@0xocdsec
原文 ↗
同类自主攻击框架已实战得手,可直接对照自建 Agent 权限边界。
Extensive and good investigation by @PandaRE__ on a Chinese actor running AI powered agents in attacks against European telco infrastructure. And it seems they had success.
ai_agentllm事件报告
@NetAskari
原文 ↗
完整自主 AI 攻击 harness 的原始工作区泄露,含手法与配置细节。
RT @PandaRE__: New blog drop! 😊
Operation Snowball 🇨🇳
While sweeping open directories on @etugenio, I found a workspace of an autonomous AI intrusion harness. The TA used a jailbroken DeepSeek agent and ran two weeks of attacks against a European telecom provider:
https://t.co/rL7DcyPVu0
ai_agentllm越狱事件
@domchell
原文 ↗
桌面端提权即拿到完整控制权,PoC 可直接验证环境是否暴露。
RT @gergely_kalman: New macOS LPE https://t.co/H6H0PyME0y
lpemacospoc
@0xocdsec
原文 ↗
0day 尚未修补,LLM 独立跑完从找洞到利用的全流程。
RT @CrackmesOne: I got two full guest-to-host escape for VirtualBox and UTM on the latest version! I used GPT-6 Sol with daybreak blue to find the vulnerabilities (0day as of now), and then used open models to build the full exploit chain. I am surprised that I can make VM escape full chain now https://t.co/WBV2K8CWpD
逃逸虚拟化ai_agent0day
@0xocdsec
原文 ↗
首例方形天线整机攻破,说明硬件层仍是卫星终端的最后短板,并附有可复现的读取演示。
RT @DarkNavyOrg: We just got a root shell on a @Starlink terminal antenna! To our knowledge, this is the first full exploit of a "square dish" since @LennertWo's attack on the "circular" one in 2021.
@SpaceX security team has done an amazing job, leaving us no choice but a *hardware attack* to achieve this 👏 The demo below shows what's only possible on actual rooted hardware: reading fuse register contents, accessing hardware-encrypted "file_edr" data, and more.
刷新公开漏洞奖金纪录,反映浏览器高危利用链仍是最值钱的一类目标。
RT @rebane2001: my chrome vulnerability was just awarded the highest bounty in @GoogleVRP history - $1,333,337 https://t.co/ePBNS5Bsr3
推荐15
低成本硬件加开源软件即可复现的无人机侦测方案,还在扩展 RID 之外的信号。
RT @cemaxecuter: WarDragon getting pushed hard! 13+ mile drone detections and tracking on things besides RID…Years of hard work poured into this platform, and it just keeps getting better and better. 🐉📡 https://t.co/cM0C3nF8Gl
公开研究多停在PoC级,此议题补上可利用性评估的判断标准
At OffensiveCon Tokyo, Nils Ole Timm (@Firzen14) gave an excellent talk about how he and other professionals assess whether bugs are suitable for production-grade exploits. Rather than the sort of POC-grade exploits that you see in public research, conference talks, etc.
It was a look behind the curtain that is quite rarely available to young up and coming hackers
Perhaps they will publicize at some point
可在消费级硬件本地运行的后训练Qwen模型,适合做Agent基座
RT @atomic_chat_hq: Introducing Nous Research Arete-27B GGUFs, post trained QWEN model.
https://t.co/sHP239g0mS
内核与 TXM 补丁细节公开,可作为 iOS 27 攻击面研究起点。
RT @Little_34306: doublera1n: a jb iOS 27 for iPhone 11 series use usbliter8 is now out!
This is different w/ CFW becuz it use the FS volume as a shared but on another Data volume with kernel, TXM patches, which reduce used storage like dualboot
PR, issues are welcome!
https://t.co/8Qb81Wg2Qw https://t.co/4OapiuKywI
后渗透规避能力清单,可对照现有检测规则的覆盖缺口。
Some features i have implemented for the past months for Vanguard Framework ..
- CET Support.
- Custom frames chains + thread entry point options.
- stack spoof capabilities during sync & async bof executions.
- evasive .NET + PowerShell executions...
- more custom post exploitation commands for engagements.
Also some cool features:-
- Rust VM = ) ?
- various dumping methods ?!
- cs bof support ?
- Multiple Extensions support ?!
- Multiple listeners support ?!
There are a lot of capabilities which I've built and integrated and they'll be explained in detail in an upcoming blog post which will be out in an couple of weeks after we got the license permit..
if you think any new ideas or features or questions ? feel free to reply & reach out...
I want to gather some insights into the features people would be interested in for modern adversary simulation and red teaming = )
给出用虚拟化隔离运行 shellcode 的具体实现,步骤可复现。
Executing Stardust Shellcode in an Isolated vCPU
New Medium post, this one is the next part of this post:
First part (Executing Shellcode in an Isolated vCPU):
https://t.co/h9QnnkAdD5
Second (new) part:
https://t.co/TgtITDrq9M https://t.co/i3meusxbIx
shellcode虚拟化poc
@papadelospapas_
原文 ↗
影响 Exynos 机型锁定引导链,属厂商已编号的可利用提权面,含技术说明。
RT @LethalAccess: Run unsigned code in Download Mode on Samsung (run custom images on locked bootloader for Exynos) - SVE-2026-3546 (CVE-2026-21155)
https://t.co/WUE82E95JY
移动端固件提权cve
@alkalinesec
原文 ↗
同时给出可复现的攻击步骤与蓝队可见的日志断点,攻防两侧都能直接对上号。
''php2shell in the Lab, and Where a SOC Catches It''
#infosec #pentest #redteam #blueteam
https://t.co/6bvDFC7hZU
rceweb蓝队报告
@CyberWarship
原文 ↗
从基础原语讲到 C2 实现,是理解与检测自定义植入体的少见系统性材料。
RT @r1cksec: An introduction to Linux internals for building C2 agents from scratch.
https://t.co/1w86bNzegJ
#infosec #cybersecurity #redteam #pentest #linux
c2linux内核议题
@0xocdsec
原文 ↗
本地进程通信的通用拦截工具,适合审计特权守护进程的攻击面。
RT @andrd3v: intercept every XPC message on macOS. any process, sandboxed daemons included
https://t.co/PAvRzAVmoG https://t.co/cklXqyXYbW
macos工具ipc
@0xocdsec
原文 ↗
被动与半主动侦察、关联与风险打分一体,可直接部署试用。
Obscura — Threat Hunter & Attack Surface Management Platform A single static Go binary for passive & semi-offensive reconnaissance, correlation, risk scoring, and continuous monitoring. https://t.co/NfmRiTC2qq https://t.co/U4kmSmLVz4
附带利用路径的完整 writeup,可照做复现验证。
RT @unsignedbytes: Writeup https://t.co/DIAuDYx44x
虚拟化边界被突破,自托管宿主需提前评估隔离与补丁策略。
RT @killa: welcome to my hyper-v escape tutorial, please like, comment and subscribe!
(s/o to @bet3rd for editing this 41 minute clip and @calif_io for the shigure ui inspo. yes its been shipped to MSRC and no, i won't tell you how to do it)
xoxo https://t.co/WfUuR6sfBs
逃逸虚拟化hyper-v
@0xfluxsec
原文 ↗
在无前沿模型权限或预算受限时,提示词设计仍显著影响漏洞发现成本。
RT @0xAsm0d3us: 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漏洞挖掘
@0xAsm0d3us
原文 ↗
将宿主进程内存映射进 guest vCPU 并代理 API 调用,是可复用的研究用开源底座。
HyperLoader — Hypervisor isolated shellcode execution environment. Maps host process memory into a guest vCPU, intercepts DLL execute faults, and proxies API calls to the host, Stardust PIC shellcode compatible https://t.co/XFK6m8Amij https://t.co/x9vIh08IgL
更多52
关系到可本地部署 27B 模型的实际规格、量化占用与推理速度预期。
RT @teortaxesTex: This is becoming interesting
- let's say Qoduck is Qwen4-27B
- I'll accept that it's ≈35GB in Q8
- 3.8-Flash-next was arch preview for 4
- Qwen 3.x-27B in fp8/q8 is ≈29 GB
- Engram is also 8bit, so 6Gb, ratio checks out
- it goes at 480 tps
- it's an extreme yapper. Qwen's old 35B 3AB was also a yapper.
Now, DeepSeek's original Engram paper used a 27B 4AB MoE (their default test shape)… I'm not sure how Engram plays with dense models or if Alibaba would be willing to check…
in short, what I'm saying is that Qwen4-27B may be a sparse MoE, unlike all previous Qwens-27B. At the level of capability we're seeing, this would be kind of astonishing and unlocking a lot of use cases. But unsurprising (3.8-Flash-next is 125B 6AB + 51B n-gram embedding and 4B MTP)
I will assign this very little probability, this is more of a wish
来自模型厂商的能力方向表态,可作 agent 自主性上限的参照。
RT @gdb: frontier models, with lots of test time compute, are becoming the engine of scientific discovery
llmai_agent
@0xocdsec
原文 ↗
浏览器代理是本地 agent 最常见的落地形态,直接牵涉其权限边界。
RT @raealisa: watching AI navigate the browser for me https://t.co/XJ3o0vbMhi
ai_agent浏览器
@0xocdsec
原文 ↗
反映检测工程中的真实疑问,但无上下文与结论。
@cyb3rops What does this mean for the effectiveness of static signatures?
澄清取证厂商能力传闻,避免据此误判移动端威胁模型
Well, that report was just not really factual it seems. The 'GraphenOS' was a self-made fork of a reseller that lacked serious security audits: https://t.co/Bij1flf7zo
提醒内网已被突破时Agent蜂群救不了场,应先削减暴露面
RT @HackingLZ: All the offense vs. defense stuff gets framed wrong anyway. At the end of the day, what’s the attacker’s motivation? If it’s financial, which it often is, how are they going to monetize the attack?
Work backward from there. Take unnecessary services off the internet, properly segment your DMZ, put things behind Tailscale or an application proxy, and reduce the attack surface in the first place.
If they’re already inside your network, launching a swarm of agents doesn’t magically remove the constraints of their initial foothold. They’re still potentially bottlenecked by a single compromised endpoint, its privileges, and whatever network access it has. Scaling up activity also risks creating a bunch of noisy C2 channels and generating significantly more alerts.
That creates opportunities for defenders to respond quickly, or ideally contain the activity automatically with a low false positive rate. A swarm generating an insane amount of suspicious activity might actually be easier to detect than the slower, more deliberate copy and paste playbooks many attackers have traditionally relied on.
ai_agent攻击面议题
@0xTriboulet
原文 ↗
若属实为迄今最大规模数据销毁事件,需跟进核实与手法细节
RT @cpartisans: Destroying 1.8PiB of data in a cyberattack breaks any world records? asking for a friend.
@ROSSTAT hack
#RosstatLeak https://t.co/jLj2SN19T0
事件wiper数据破坏
@0xocdsec
原文 ↗
RT @Sandbagger_01: UK sends cyber ‘hunt forces’ abroad to stop hostile hackers
https://t.co/loAhUGy54N
RT @brentius_: The chip in question...
And yes, it does beat a H100. By almost twice. While using 160x less power. On one very niche and specific workload. But it's a win, and I'll take it. https://t.co/ezLvdWJYDp
RT @uzairakrum: Me showing my abliterated Qwen 4 27b to chatgpt and Claude users next week https://t.co/h3xHjJHbnk
In my own head, I have a sort of visual-spacial representation how how the various system internals of iOS work
Many other vulnerability researchers have said that their minds work similarly to reason about systems
So what I did was try to create 3D representations to map functions across all the different software/firmware/hardware layers of abstraction that these subsystems work (in this case, iOS 27's SPTM). One of the 3D representations even presents an immersive 'video game' of sorts, which lets the agents "walk through" the system
I have found this to be a helpful resource for agents, keeping them more tightly grounded to the functionality of these complex systems, though I don't have formal benchmarks (which are overrated anyway)
If Meta were to open source Muse Wonder, they would absolutely put themselves back on the map of mainstream AI power-users
The ecosystem would improve on their model, and give them ideas for how to do so themselves
Even though my first impressions were not so great, the model is currently working spectacularly for certain use-cases
218,070 donations at https://t.co/mJ2BhqyJo7, I am incredibly thankful for all of that; there's thousands coming in per hour. I will kick off a massive validation run in a bit, retrain the classifier and then figure out how to distribute the whole dataset scalably.
RT @ptrpaws: android persistent threat https://t.co/zyLN4Zggjq
攻击面来自第三方 fork 与远程管理模块,而非官方系统本身。
RT @T3chFalcon: “GrapheneOS cracked by police”
The source article for that claim says there’s no sign of an exploit and no sign the phones were running official GrapheneOS.
the dutch phones came from a reseller who built his own fork, reportedly with AI tools and added remote device management so he could control the phones himself.
That’s a stranger’s software with Grapheneos in the name and a remote control built in not GrapheneOS.
About the “cracked” german phone, the police seized a Pixel 8a in 2025. It failed extraction that july. A year later they got data off it. A seized phone can’t update, so it was still on its 2025 patches the whole time. How they got in is undisclosed.
自托管 agent 若沿用同类拦截策略,会波及合法安全工具链。
RT @Ch0pin: @AnthropicAI , your "cyber classifiers" make your models useless even for benign programming tasks. It's not a cyber crime to create a decompiler... really doesn't make sense what you are doing.
llmai_agent绕过
@0xTriboulet
原文 ↗
RT @GrapheneOS: @realNyarime No, the @mano_crypto account is posting fabrications not matching the sources. The original article is at https://t.co/oKvOivCY2K and you should read it for yourself. It says there's no indication of any exploit and no indication it involves official GrapheneOS releases either.
RT @fuelkek: Russia raided nearly 100 suspected crypto criminals.
What investigators found inside their properties was absolutely insane.
Helicopters. Rolls-Royces. Bentleys. A Tesla Cybertruck. Mountains of cash.
All connected to a $1.4 BILLION crypto laundering operation.
Here's the story. 🧵 https://t.co/18BXcuyVPI
You guys aren't going to believe this, but.... The only model that comes close to GPT 6 Daybreak for actually writing advanced exploits is....
Muse Wonder
(but only once you really give it proper context)
RT @tomshardware: Pocket-sized Ukrainian receiver intercepts Russian strike drone video feeds up to six miles away — device only slightly heavier than a smartphone, intercepts analog video signals sent to Russian operators https://t.co/rQLoqAvIq3
RT @IntCyberDigest: ‼️ A British soldier kept his SIM card in on a NATO deployment near Russia, against orders. Someone hacked his phone, took his nude photos and blackmailed him.
He paid £100. When a second demand came, the corporal reported it himself, then was court-martialled for disobeying the order.
He got 10 months' military detention and was demoted to lance corporal. Troops in Tapa, Estonia, had been told again and again to pull their SIMs.
Appeal judges upheld the sentence. No proof of hostile spies was needed, they said, as an unknown attacker getting into his phone showed the risk was real.
Also its super cool when i worked with Control flow Enforcement where i have faced several challenges during the post exploitation parts..
CET is not new and few commercial framework has already implemented it in their own ways..
RT @NSTRIKE1231: 🇺🇦📡‼️ UKRAINIAN COMPANY MILITECH GROUP DEVELOPS NEW LTE MODEM TO KEEP DRONES OPERATIONAL EVEN WHEN STARLINK ACCESS IS RESTRICTED!
Ukraine’s Militech Group has introduced GRIM, a domestically developed LTE modem designed to maintain reliable drone communications when Starlink connectivity is limited or unavailable.
According to the company, the device can amplify LTE signals by up to six times, helping maintain stable connections even in areas with weak cellular coverage.
GRIM has reportedly completed testing and is ready for serial production.
Militech Group representative Vasyl Polishchuk explained the importance of alternative communication systems.
“Considering the disadvantages of fiber-optic cables, including cost, range limitations and unreliable suppliers, as well as the problems associated with Starlink, LTE is becoming almost the only truly reliable solution for unmanned systems.”
🔥 The development highlights Ukraine’s efforts to diversify drone communication technologies and reduce dependence on any single provider.
Reliable alternative connections could improve operational flexibility and help Ukrainian forces maintain drone capabilities when satellite communications are disrupted or restricted.
See the latest updates with us: @NSTRIKE1231
针对 Agent 谎报任务完成的失效模式引入自动验证闭环,与本地 Agent 可靠性设计直接相关。
RT @thesupermannx: China cracked recursive self-improvement 🤯
Tencent dropped 27B model that rewrites, retrains, and refines itself.. with zero human in the loop.
It’s called "Terminal-RSI" and it solves one of the biggest problems with AI.
They are incredibly lazy.
Usually, an agent will make partial progress, claim it succeeded, and completely fail the final environment verifier check. this new pipeline forces agents to actually finish the job.
here’s how it works:
they built a recursive loop with two alternating phases..
- RSR (recursive self-rewrite): the model fine-tunes itself on its own successfully rewritten trajectories
- RPR (recursive policy-refine): it then runs reinforcement learning (ppo) using a "critic" inherited directly from the previous stage to lock in those discovered behaviors
- Loop: the training sequence leapfrogs itself: Base → RSR-1 → RPR-1 → RSR-2 → RPR-2
and the numbers are insane..
they applied this to a qwen3.8-27b base model. across the four "terminal-bench" suites, the agent's average success rate jumped from 20.68% all the way to 50.58%.
on the absolute hardest benchmark suite (terminal-bench 4.0), the base model scored a flat 0.00%.
after terminal-rsi?
it hit 19.84% (putting it right on par with reference models like gemini 3.8 flash).
the secret sauce? they fixed the reward system.
instead of giving the ai "dense rewards" for just trying or passing semantic checks, they switched to a "success-prioritized reward" (spr).
it mathematically separates strict success from partial credit.
basically, the agent only gets the massive reward payout if the code actually passes the final terminal gate.. so it stops faking success.
the gap between partial local repairs and strict terminal verification is closing fast.
we are watching agents build their own training environments and level up autonomously.. the future is wild.
ai_agentllm工具
@0xocdsec
原文 ↗
本地跑模型时定位显存占用来源的开源小工具,可即刻上手。
RT @tomshardware: Open-source GPU task manager reveals exactly what's eating your VRAM — Microsoft VP's new GPU Taskman tool reveals every process using your GPU, how much memory each uses, and breaks down AI and LLM workloads https://t.co/ikzOnWrp3P
工具llm
@artem_i_baranov
原文 ↗
RT @yo_yo_yo_jbo: So, unfortunately my CTF challenges won't be available at OBTSv9, so I'm releasing them now, publicly!
Note that those were coded (by hand!) 7 months ago, so the trouble Opus had with solving them might not be correct anymore.
Give it a try!
https://t.co/gq15wKrW0W https://t.co/YU5WzYtQXH
everyone please submit your challenges we need challenges thanks
RT @mvanhorn: Not enough people are talking about how big of a deal Reverse Engineer Anything is. Peep this stat, #1 software project on all of of GitHub with 442k stars but also 429k stars IN THE LAST MONTH. What will you build?! https://t.co/vKRwL6lnS5
RT @vincent_koc: My new setup for REA skill... Grok Super + GLM 5.2 Abliterated + Deepseek Flash Heretic. 🕵️
RT @thomasmc: Life hack if your flight doesn't have Starlink: ask your agent to bypass the 443 split-TCP proxy via a $0.03/h BBR multiplexer and get ~4× faster, reset-free internet.
Open-sourcing it from the clouds: https://t.co/TBH5uRX9AO https://t.co/zeGQv0PfJA
RT @vishalsingh2972: Qwen 4 27B almost beating Opus 5.5!
who would've thought? 😱 https://t.co/AzZLNxJijC
把 LLM 接进情报流水线的可复核实现,各步骤保留引用来源。
RT @Dinosn: Argus - AI-powered adversary intelligence for purple teams. investigates a threat across the web, your RSS feeds and documents you supply, extracts quote-backed claims with an LLM, consolidates them into confidence-scored assessments mapped to MITRE ATT&CK, and lets a human review every step. https://t.co/RN1dmBgUIB
ai_agentllm工具
@artem_i_baranov
原文 ↗
带 PoC 的新漏洞细节已落地,可据文章自查受影响版本与配置。
RT @Licht_jb: A quick write-up and PoC.
https://t.co/MihZYNiAEb
cvepoc
@patrickwardle
原文 ↗
本地可跑的多格式样本快速分诊,省去逐个工具串联的时间。
FlatScan — Static malware triage engine in Go: hashes files, extracts strings and IOCs, parses PE/ELF/Mach-O/APK, scores capabilities, and emits JSON, YARA, Sigma, STIX and PDF reports. https://t.co/i1bc2jRvaa https://t.co/qVIvaW5J3C
工具malwareyara
@KitPloit
原文 ↗
内网横向与硬件攻击的实操速查,覆盖常规载荷库之外的场景。
Many of you used 𝗣𝗮𝘆𝗹𝗼𝗮𝗱𝘀𝗔𝗹𝗹𝗧𝗵𝗲𝗧𝗵𝗶𝗻𝗴𝘀, but do you know about the rest of the "𝗔𝗹𝗹𝗧𝗵𝗲𝗧𝗵𝗶𝗻𝗴𝘀" family? 🧐
- 𝗜𝗻𝘁𝗲𝗿𝗻𝗮𝗹𝗔𝗹𝗹𝗧𝗵𝗲𝗧𝗵𝗶𝗻𝗴𝘀: Active Directory & Internal Pentest
- 𝗛𝗮𝗿𝗱𝘄𝗮𝗿𝗲𝗔𝗹𝗹𝗧𝗵𝗲𝗧𝗵𝗶𝗻𝗴𝘀: Hardware & IOT
🧵⬇️
工具内网iot
@pentest_swissky
原文 ↗
常用载荷库改为在线可查,检索与贡献成本都降低。
As always feel free to make a Pull Request to add your knowledge🤗
PS: PayloadsAllTheThings also has its own dedicated live version https://t.co/9S5AkbcP9m
工具payload
@pentest_swissky
原文 ↗
硬件与 IoT 测试的入口级资料,含固件与协议页可直接上手。
𝗛𝗮𝗿𝗱𝘄𝗮𝗿𝗲𝗔𝗹𝗹𝗧𝗵𝗲𝗧𝗵𝗶𝗻𝗴𝘀 Your go-to cheatsheet for everything related to Hardware and IOT. You will find a lot of resources about gadgets such as FlipperZero, HydraBus. Also have a look on the firmwares and protocol pages 🤓
https://t.co/yjv9t905Dx
工具固件iot
@pentest_swissky
原文 ↗
自托管大模型的硬件门槛参考,非安全问题。
RT @0xKevin00: 把一台迈巴赫放桌上是什么感觉?
用装机视角看本地部署 768G 显存跑DeepSeek V4.1 🤓 https://t.co/uenkm9AhU4
AI 厂商第三方数据接触面的争议,细节尚未核实。
This security company cathedral basically breached Anthropic saved customer PII (idk why) and then gave a pinky promise that they deleted it? am i misunderstanding something here?
移动设备跑替代内核的实验进展,与安全无直接关系。
RT @Dan50798734: ReactOS booted on the OnePlus 6 (Snapdragon 845) https://t.co/eJaTDRExv6
Linux 内核态反作弊的结构性弱点,绕过思路可迁移到其他反作弊。
RT @makemake_kbo: Bypassing EACv2 readiness checks. Ready for day 1 🫡
Will share source code eventually, essentially unpatchable with their approach to AC on linux
@TimSweeneyEpic my dms are open 😘 https://t.co/wU40SXuQGW
he is also a ctf player and its crazy to me how much effort he puts into his blogs https://t.co/lQp7XH0rUV
身份伪造已降到消费级成本,视频面试与社工诈骗可复现。
RT @Oddvarmoe: @IgnoulJoris Deep Live Cam for face https://t.co/omsyadN9XL and also used Osb studio and a virtual camera.
For voice training and realtime changing I used https://t.co/sj8uknrSgX also used a virtual cable driver to redirect sound
沙箱是最后一道隔离边界,案例可对照自建浏览与渲染环境。
RT @davi1337_: I wrote an article about browser's sandbox bugs
https://t.co/JbGnVLfNhR
把 DNS 管理权交给 Agent 即扩大攻击面,接口与鉴权设计值得先看。
Windows-DNS-Policy-Manager — Browser-based GUI for Windows Server DNS — policies, zones, records, DNSSEC, DNS over HTTPS, and blocklists. Next.js frontend + PowerShell bridge, 8 scenario wizards, and an MCP server for AI agents... https://t.co/BWlBS9CNPr https://t.co/AMjBwyhVng
工具mcpai_agentdns
@KitPloit
原文 ↗
universal content policy bypass
im not gonna share it
一条泄露凭据即可跨租户横向,云权限模型需按此复核。
RT @SpecterOps: A vendor holds credentials. A service principal holds permissions in your tenant.
For multi-tenant apps, a compromised cred can expose multiple customer environments.
Read more on app registrations, service principals & consent connect from @rbnroot ⤵️ https://t.co/dlQjDSj3wg
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
前沿实验室如何获取安全数据,直接决定攻防两端模型的能力落差。
RT @thedealdirector: OpenAI and Anthropic have a tabula rasa problem with cybersecurity.
It’s been barely six months since the frontier labs started investing heavily in security data for model training. Once Mythos landed, it was clear the “AI breaks the game” moment had arrived.
Anthropic claimed to be very focused on working with the cybersecurity ecosystem to “integrate the model safely”, but in reality it ignored almost every conventional approach in the industry.
The biggest problem is the very one-dimensional narrative the frontier labs are building, in which the technical side of AppSec is treated as the whole field of cybersecurity.
Security is first and foremost about people, process and systems. Technology plays a useful role, but the areas where we try to reduce risk are numerous and hard to tackle. Finding and fixing bugs is just one part of the game: sometimes very important, sometimes completely irrelevant.
Exploiting a zero-day in a critical system doesn’t achieve much if the attacker has no way to move laterally within the network and only limited access to relevant data. There’s a reason the majority of successful breaches come down to insider risk and stolen credentials rather than the exploitation of novel bugs.
What AI has clearly democratized is access to zero-days. They mostly used to be in the hands of a far larger group than the nation-state actors, elite ransomware crews and occasional 200 IQ obsessive loner who traditionally had them. Today, they’re available to anyone who can spin up an abliterated LLM.
What it has really exposed is how poorly almost every company manages risk across people, process and systems. Rather than working with the wider security community to reshape how we work, the frontier labs have often chosen to be adversarial with key researchers and practitioners in the field.
To some extent, their goal is to replace every security vendor. That can look greedy, but it’s arguably well deserved. Compared with most other software categories, security vendors are far more likely to:
- Sell you software that doesn’t work at all
- Sell you software that works but breaks when integrated with the rest of your tools, making it useless
- Sell you software that works and fits into your workflow but comes with predatory pricing and deliberate lock-in
It’s one of the griftiest corners of the software industry, and the majority of practitioners hate almost every vendor with a passion. Those vendors have responded by pouring money into distribution, building large sales organizations that burn a lot of cash for very little in return.
So the frontier labs looked at this and decided to create a new problem, become the main solution to it, and then take over more and more of the industry’s technology layer over time.
Pretending that everything after Mythos is a blank slate, and that we’re only now “discovering” cybersecurity, is cynical at best and highly irresponsible at worst.
My message is simple: this strategy will fail unless you structurally focus on people and process first. Whether the leadership at OpenAI and Anthropic cares is another matter. They’ve shown repeatedly that they don’t care about the opinions of players in the ecosystem unless they’re big API consumers.
But don’t try to erase 30 years of ideas, practices and technology just because they get in the way of token consumption.
llm供应链报告
@0xTriboulet
原文 ↗
nuclei — Updated! — Fast YAML-based vulnerability scanner with template-driven detection engine for automated security testing across web apps, APIs, networks, DNS, and cloud infrastructure. https://t.co/EYqqAkugbQ https://t.co/MYsD4kSapv
Well this is genuinely crazy. Using VMs don’t save you anymore 💀
crazy work congrats @killa
> supply chain compromises and GitHub AI pentest projects will have the biggest impact
100%. It’s just not the sexy thing to say.