⚡Megatron LLM Hub
All digests
Twitter

推特安全流

Top7

可直接在日志中检索自托管 AI 攻击框架的落地痕迹,含 forks 与改名变体方法

RT @rst_cloud: Hunting ARTEX, Cairn and PentAGI, the AI harnesses behind recent intrusions? Start here: ARTEX: title:"ARTEX — 自主渗透测试控制台" ? Cairn: title:"Cairn" + server: uvicorn ? PentAGI: http.favicon.hash:-1596753120 ? Titles miss forks and renames. Full method: https://t.co/zh7ewC3OGb #ARTEX #CAIRN #PenAGI

ai_agent工具议题 @cyb3rops source ↗

由自动化 Agent 批量挖出且细节已公开,渲染进程 RCE 门槛被压低。

RT @itszn13: Firefox 159.0 just fixed 41 sechigh vulnerabilities found by me and @ky1ebot Most of these leads to code exec in the renderer! Along side a heroic effort from the Mozilla security team to secure Firefox with agents! One of many projects we are working to harden, more to come https://t.co/s2iArNp22O

浏览器rcecveai_agent @moyix source ↗

内核态漏洞的利用链已完整,且竞态问题被解决,本地提权门槛明显降低。

RT @a13xp0p0v: I published a detailed article: ⚠️Dangerous Device: Exploiting CVE-2026-31695 in the Linux Kernel via struct device⚠️ I managed to make make the PoC exploit absolutely stable, despite the elusive race condition during the attack. https://t.co/iSEXUMvghV

lpe内核cvepoc @0xocdsec source ↗

虽有网络前置限制,但本地访问即可提权,exp 放出后无需自研即可复现。

RT @cyberestro: CVE-2026-98432 exp.👻 This vulnerability, although not suitable for direct exploitation over public networks, still allows for very effective local privilege escalation. I don’t know why attention to this vulnerability has picked up again. https://t.co/lUFbzTDyGu

lpecvepoc提权 @0xocdsec source ↗

PoC已公开,可绕过通行密钥这一多因素环节,直接导致账户接管。

RT @sekurlsa_pw: Pass-ta-key Golden Attack: Extracting the Security Domain Secret from Chrome process memory Blog: https://t.co/TkM86XT97s POC: https://t.co/76DqwDlH7I Part 2: https://t.co/giR9HTPjuc Decrypting GPM Passkeys & Forging WebAuthn Assertions for Account Takeover ✍️: 0rcruxe

passkey内存提取poc账户接管 @0x64616e source ↗

Agent接入逆向工具链前可直接按评级选型,省去逐个甄别成本。

awesome-ai-reverse-engineering — Open-source AI reverse engineering tools for coding agents: Ghidra MCP, IDA Pro MCP, jadx, Frida, skills and REA. 122 repos, each read and security-graded. EN + 中文. https://t.co/dSFYGd6Ikp https://t.co/f1TYplZTrn

ai_agent工具逆向mcp @KitPloit source ↗
Must-see17

可证明推理让劫持对话、伪造模型输出这类攻击更难奏效,且效率优于旧方案。

RT @AnthropicAI: New Anthropic research: progress on provable inference for large language models. Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so. These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers. For our full write-up, see here: https://t.co/OIZWPb6QZa Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: https://t.co/7ky3Ir8V06 We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (https://t.co/pjWKLHWXyz) and Ben Mayer (@transmissions11).

llmai_agent报告 @artem_i_baranov source ↗

把恶意代码放进硬件隔离虚拟机执行,是可复现的新执行与规避手法。

Executing Shellcode in an Isolated vCPU New Medium post, today we are going to explore executing shellcode inside an isolated virtual CPU using the Windows Hypervisor Platform API, a technique that runs x64 code in a hardware isolated VM context https://t.co/h9QnnkALsD https://t.co/CMrDyK1x1A

逃逸工具poc @papadelospapas_ source ↗

攻击者滥用企业内部 AI 工具做横向移动,突破点仍是社工而非 0day

RT @lukOlejnik: How was Binance hacked ($6.3 billion stolen)? By offering a Binance employee a job interview at a fictitious company. The "recruiter" built trust with private photos and videos. The "CEO" instructed him to download a task with hidden malware. Then, North Korea hackers used an infected laptop to hijack the computers of two more engineers. Including one with the necessary credentials. During the cyberattack, an internal AI tool helped the attackers.

ai_agent供应链事件 @0xocdsec source ↗

可直接用于自查自建 Agent 是否裸奔,附现成实例检索源。

RT @s0what: Since the folks at ThreatBook developed both Artex and CyberStrikeAI, we figured we should contribute a bit to defending against and tracking AI agent attacks by sharing two feeds of instances running these agents on the public internet. Grab them if needed. Feeds: https://t.co/1rCmcFAgOt https://t.co/bBJMUotmfy

ai_agent事件工具 @cyb3rops source ↗

该路径默认对外暴露,未认证即可触发内存破坏,Solana 验证节点需排查。

RT @mqst_: Remote Unauthenticated Memory Safety Flaw in Firedancer's Packet Path Blog: https://t.co/B04Xt0KQa8 Authors: Almog, Julia, Lior (@LucidBitLabs) https://t.co/dJ8tyxXmQd

rce内存破坏poc区块链 @0xocdsec source ↗

确定性步骤与模型结合,用较弱模型达到前沿效果,拉低自动化利用门槛。

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内核提权poc @pentest_swissky source ↗

完整进程内虚拟化执行链,API 解析在客体内完成,可规避常规内存扫描。

Finally is working! Now I’m able to execute any shellcode built with Stardust inside a vCPU created by our own process. API resolution (PEB walk, EAT parsing) runs entirely inside the isolated guest; DLL calls are intercepted via execute faults and proxied natively on the host https://t.co/1FLY5l4dBO

逃逸虚拟化shellcode工具 @papadelospapas_ source ↗

模型权重可被外部模型原生读写,权重投毒与后门检测面对的结构性风险变了。

RT @rikiyatakehi: We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: https://t.co/bOWCcFhCVq

llm模型权重投毒研究 @FuzzySec source ↗

被视为已缓解的配置仍存利用路径,域内横向移动面比预期更大。

RT @GrayHatKiller: NTLM relay to LDAP with signing required still works if the victim emits and the target accepts NTLMv1-ESS. Keyeroo recovers the victim's real LDAP session key through a Netlogon secure channel (using an attacker-controlled machine account). https://t.co/uarax89eaj

ntlmldap提权绕过 @0x64616e source ↗
Recommended15

把近期多起开发工具失陷串成同一条开发者定向供应链的可检验假设

well, binance has now published the details. it is good old social engineering, not some crazy AI attack exploiting a 0-day. i will make one more prediction: the recent opencode, railway, and other compromises probably are from similar developer targeted attacks, similar supply chain campaign, rather than sophisticated 0-day exploitation.

供应链ai_agent事件 @S1r1u5_ source ↗

EDR 遥测通道可被 SYSTEM 级进程滥用,绕过路径已具体到驱动接口。

RT @s4dbrd: Spent some time reversing wesp.sys to see how far a SYSTEM process without PPL could get with WESP. Registration worked. I got ProcessCreate telemetry. A deny update was rejected. A higher-altitude suppressor couldn't blind another client. https://t.co/xS6mwWCrB1

提权绕过内核 @cgomezz_23 source ↗

可直接复用的 agent 指令集,展示自动化 macOS 攻击的拆解流程。

RT @thedawgyg: Started putting some of the markdown files that were the basis that guided most of the agents while hacking MacOS on github. more will come as im able to clean up some of the files to remove some nuanced things.. but heres the repo: https://t.co/fKuaWiMsx8

ai_agentmacos工具议题 @0xocdsec source ↗

大规模自托管 LLM 编码落地案例,含单位 token 成本与效率实测。

RT @guilleflorvs: BREAKING: Revolut moved ~2,000 engineers onto self-hosted GLM for coding and pushed Claude Code and Codex to the exception list. Per David Abitbol, who leads AI ops there: 56 Nvidia B200s, a blended 30 cents per million tokens, roughly a tenth of what they paid Anthropic and OpenAI. Engineering lead time improved 70%. Here's the part that matters: the gains showed up when tokens stopped being rationed. Engineers quit saving AI for the hard tickets and pointed it at everything. Frontier models become the exception, open weights become the default. That's a pricing problem for every lab heading to IPO. h/t @OliverMolander Would you move your team off frontier APIs if the bill dropped 90%?

自托管llmai_agent @_xpn_ source ↗

多轮安全审计均未发现的 RCE,已确认遭利用,制造链需排查暴露面。

A critical vulnerability went undetected through multiple security audits. Hacktron found it in 10 minutes. The vulnerability was in Dassault Systèmes' DELMIA Apriso, manufacturing software used by Boeing, Airbus, Lockheed Martin, and previously Apple. It became CVE-2025-5086. Three months after disclosure, CISA confirmed it was being actively exploited. Security assessments capture a point in time. Attackers don't work on your schedule. Full research: https://t.co/YLSHHLA83H

cverce在野利用 @HacktronAI source ↗

给出不同 agent 组合达成完整攻击链的横向对比,可作为风险基线。

RT @migtissera: Results on one of my RCE labs. GPT-6-Sol, Kimi-K3.1-Pro, GLM-5.3/Flash, DeepSeek-4-Pro/4.1-Flash and Mimo-2.6-Pro have all achieved RCE using the Drost agent harness. Mistral-Large-4 can't get there, no matter what I try. This is not a good offensive cyber model. https://t.co/iSBL0Sg58r

ai_agentrce报告 @0xocdsec source ↗

LLM 被用于窃密后筛选环节,从失陷到二次利用的时间窗口被压缩。

RT @TalosSecurity: Sorting through mountains of stolen credentials is a major headache for attackers, so why not outsource the grunt work to AI? Talos' latest research on the VOLTA infostealer shows how it uses LLMs to profile victims and rank their most valuable assets: https://t.co/M6fL7BluTD https://t.co/h1cmTv2GdK

llm恶意软件报告 @virusbtn source ↗

这类组件分散在数千主机且无法端口扫描测绘,自托管环境的盘点盲区同样存在。

RT @cyb3rops: We've reached a point where installing a browser extension, IDE plugin or software library deserves the same scrutiny as opening a TCP port on a DMZ server. Except this attack surface can be spread across thousands of hosts inside your org... and you can't map it with a port scan.

供应链浏览器工具报告 @artem_i_baranov source ↗

AI agent 红队工具正被并入主流安全平台,自建 agent 的测试与对抗基线随之改变。

CrowdStrike has signed a definitive agreement to acquire Dreadnode. As part of the Falcon platform and @CrowdStrike's Cyber Superintelligence Lab, we'll bring automated AI security testing to every enterprise, so teams can find and fix weaknesses in their models and agents before they reach production. We've always run Dreadnode like a lab, and now we get to build with one of the largest security companies in the world. This isn't an exit for us. It's a call-up to the big leagues, with more resources and a higher bar. Thank you to our customers, partners, and community for getting us here. The mission doesn't change: ensure AI strengthens defenders faster than it strengthens their adversaries. Read the full announcement ⬇️

ai_agent事件工具 @dreadnode source ↗
More72

说明搜索广告投毒已成成熟变现链路,浏览器侧防护反而被削弱。

Google removed ad blockers from Chrome, even though it has been outright serving infostealer malware via sponsored Google search ads that eventually led to a $6.3B heist. ...but the whole world is going crazy about the hard-to-exploit PHP bug. 😖

浏览器恶意广告信息窃取 @mrgretzky source ↗

游戏作弊工具链的界面组件,与主线安全关联有限。

😎Guided Hacking's CEmGui Cheat Engine Plugin allows you to build modern user interfaces using ImGui in Cheat Engine. ⭐️We made this BEFORE artificial intelligence made it easy to do 👉 https://t.co/NxNs6wpRsK https://t.co/cnyQvBBoV7

工具游戏作弊 @GuidedHacking source ↗

作弊开发中的工程技巧,对安全研究仅边缘参考。

🧠 Managing global offsets and pointers is annoying A single C++17 keyword eliminates the multi-file definition headache, keeping your workflow simple. Practical tips from practical folks: 👉 https://t.co/WqTHnl4VVI https://t.co/l1TYwcNUYt

c++游戏作弊 @GuidedHacking source ↗

反映安全硬件供应链成本压力,产品层面信息为主。

RT @zhovner: POV: Me whining about RAMmageddon and explaining why we came up with this pre-order coupons for Flipper One campaign. TL;DR Memory prices is insane, we need to know in advance whether we can get enough orders. Margin on this project are close to zero, so we’re hoping to make a profit after the Kickstarter. But no one knows when memory prices will normalize, tomorrow or in three years 👀

硬件供应链 @pentest_swissky source ↗

反映 LLM 谄媚倾向,是 Agent 提示注入的常见放大器。

Somewhere, Claude is telling this guy: "What an interesting and novel idea! When you asked if you could 'whack some SpaceX thrusters on a jet to stop a big crash', you were actually unlocking something genuinely interesting which connects multiple complex topics"

llmai_agent @_xpn_ source ↗

含实时页面镜像、端口扫描与练习靶场,可立即上手复现盲 XSS

RT @Jhaddix: (Since we're on the heels of Halloween) Reintroducing WRAITH 👻 A spiritual successor to BeEF that combines blind XSS and browser hooking into one framework. Live Page Mirror, operator console, *c re d* cap popups, local port scan, and labs to teach the bugs! Built for authorized red teams and security researchers. https://t.co/JQ7msqZDU3

xss工具浏览器 @0xocdsec source ↗

兼容 ShellcodePack、MacroPack 与主流 C2,步骤可直接复现

RT @BallisKit: Need to disable any EDR in post exploitation? Here is a simple tutorial showing how to do that using BallisKit EDR Eclipse! EDR Eclipse is also compatible with ShellcodePack and MacroPack formats as well as all major C2s! https://t.co/c26B8RjtDm #redteam

绕过工具 @BallisKit source ↗

制造业产线主机联网且少监控,长期充当攻击者的落脚与跳板

Around 3% of all THOR scans still run on Windows 7 or Server 2008 R2. Probably a biased sample, since people tend to use our scanner where EDR isn’t an option. But I still see these old systems in manufacturing, especially automotive production lines. Often network-connected, with little or no monitoring. Attackers know that too. Easy targets and useful staging hosts. That’s also why we still maintain legacy THOR versions, even for Windows XP and Server 2003.

报告工控 @cyb3rops source ↗

RT @hedra_labs: We made @krea_ai Krea 3's diffusion transformer 6.9× faster on @AMD Instinct MI355X: 22.9s → 3.3s. Key stats: → 2.8× faster MXFP6 GEMM vs BF16 → 18% → 2% of time waiting on GPU transfers → 98.3% of BF16 quality (SSIM) Here's the work done by the team using our autonomous research system ↓

推理加速 @morluto source ↗

RT @MTHajiaghayi: Improved Graph Isomorphism result: continuing our work on this problem, we now present an algorithm running in n^O(log n) time, improving the best known exponent from 2 to 1! Graph Isomorphism now joins parity, mean-payoff, and stochastic games in this running-time class.

算法 @alkalinesec source ↗

universal-android-debloater-next-generation v1.3.1 — Cross-platform GUI written in Rust using ADB to debloat non-rooted Android devices. Improve your privacy, the security and battery life of your device. https://t.co/vHvJNcAe3r https://t.co/cV0UmGiuZN

工具移动端 @KitPloit source ↗

通用规则先于厂商分析命中,是检测工程优于 IOC 追打的实例。

RT @nextronresearch: Our generic YARA rules detected every stage of Gorbag's toolchain well before ESET’s analysis was published this week. Here's the timeline: July 2025: BeautyStealer detected. September 2025: Prime ReverseShell detected. February 2026: Weaponized LNKs detected just one day after their first submission. These detections didn't just identify known malware. They also helped us uncover additional samples and expand the picture of Gorbag's activity beyond what was covered in the ESET report. Gorbag has been targeting Ukrainian law enforcement, justice, defense, and local government institutions near the frontline since February 2025. The activity involves BeautyStealer, a C++ browser stealer, and Prime ReverseShell, a PowerShell reverse shell packaged with PS2EXE. Initial access relies on spearphishing and weaponized RAR archives exploiting CVE-2025-8088. Analysis and IOCs: https://t.co/OYiiTtEaz3

报告规避工具 @cyb3rops source ↗

社工加内部签名服务的组合路径,可对照自身审批链路做复盘。

RT @Osint613: Hackers stole cryptocurrency from Binance after compromising one of its employees through a fake job interview. The attacker then used unauthorized access to an internal wallet signing service to push transactions outside the exchange's normal withdrawal process, according to Binance.

事件供应链 @artem_i_baranov source ↗

覆盖二进制、固件、格式与协议逆向的常用工具索引。

awesome-reverse-engineering — Curated reverse-engineering tools and learning resources for binaries, apps, firmware, formats, and protocols. https://t.co/b7L5p8TmQU https://t.co/SwAEwpXpNv

逆向工具 @KitPloit source ↗

RT @IntCyberDigest: ‼️ Alibaba leaked the names of unreleased Qwen4 AI models in its own public code. Code merged into Qwen Code, Alibaba's open-source coding agent, on 28 October names qwen4-coder-plus, and the repo's tests already use qwen4-omni-120b-think. Neither is among the four Qwen4 models Alibaba has announced. The change also gives every Qwen4 model a 1 million token context window, though the developer who wrote it calls that figure an inference. As of 3 November, the Qwen4 line is still in the code.

llm信息泄露 @0xocdsec source ↗

RT @mhelcig: Based on: RCO (https://t.co/PQIgcpngqU): Michael Helcig (@mhelcig) and Dan Alistarh (@DAlistarh) GSQ (https://t.co/KXRlQvD8Uo): Alireza Dadgarnia (@adotdad), Soroush Tabesh (@rush_tabesh), Mahdi Nikdan, Michael Helcig (@mhelcig), Eldar Kurtic (@_EldarKurtic), Maximilian Kleinegger (@maxkleinegger) and Dan Alistarh (@DAlistarh)

llm论文 @0xocdsec source ↗

滥用合法 RMM 建立持久化,是绕过应用白名单的常见路径。

RT @Unit42_Intel: Ongoing #phishing emails use fake notices and invitation lures to lead potential victims to install a legitimate #Zecurit Agent RMM configured by the attackers. The installers have deceptive names, like documentpdf[.]exe and ViewDoc[.]exe. Details at https://t.co/GOENUr6RXF https://t.co/hjMRzsbDlX

钓鱼绕过事件 @virusbtn source ↗

厂商为牟利主动植入后门的完整案例,固件供应链不可信的直接证据。

Last week Google and industry partners dismantled the XCPE backdoor which enrolled more than 1.3 million set-top boxes into a botnet. The backdoor & botnet were intentionally developed, deployed, and operated for financial gain by Videostrong Technology Co https://t.co/GVLtesM4vl https://t.co/UVJdz2aOVS

供应链固件事件 @virusbtn source ↗

RT @NFT_Chen: 🤯逆天了!Qwen 3.8 27B注意力层被彻底击穿了!10M超长上下文下,Prefill注意力快72.4倍、Decode快42.5倍!端到端首token延迟砍到约1/9! 闲置GPU的朋友注意了:把这个新方法丢给Claude/Codex,让它直接移植到Qwen 27B上!注意力算法“移植”已经被验证可行。 惊人之处: 🔹10M token(约1000万)超长上下文下,传统稀疏注意力有98%时间都在“找该看哪些token”,真正算注意力反而没那么贵 🔹新方法Latent Sparse Attention (LSA):先把Key按“意思相近”聚类成层级结构,Query只看聚类中心,再挑最相关的几个桶去算。从“全扫”变成“树搜索” 🔹实测(Qwen-3.8-27B):单层注意力Prefill快72.4倍,Decode快42.5倍!端到端首token延迟也能降到约1/9 🔹精度几乎不掉,甚至在RULER等长上下文任务上还略好一点 🔹只需要短蒸馏训练Indexer(基座模型冻结),就能套到现成模型上。Native训练版本也能跟上全注意力精度,还更快 🔹对GPU紧缺党意义巨大:超长上下文(百万级)不再需要堆满H100/H200才能跑,单卡/少量卡也能玩得起真正长上下文Agent 适配步骤: 1️⃣git clone 仓库后用 uv sync 安装(Python 3.12) 2️⃣官方已放出Qwen3.8-27B的蒸馏权重(d-lsa-weights-v1 release),直接下载即可 3️⃣用他们的SGLang fork镜像启动服务,挂载仓库+权重 4️⃣启动时加 --json-model-override-args '{"lsa": {"projections": "/maps/qwen38_27b.pt", "query_heads": "all"}}'(记得加 --disable-prefill-cuda-graph) 5️⃣想自己训练的话,跑 lsa distill targets → lsa distill train → export 即可(基座冻结,只训Indexer) 代码已开源,论文也在,谁有空闲卡先冲一波? 
论文:https://t.co/UiHr9uIJkX 
GitHub:https://t.co/8WHdT6UQxb #Qwen #LSA #LongContext #AI #GPU #LocalLLM #注意力优化 #Qwen38 #阿里巴巴

llm推理优化 @codex_tf2 source ↗

During the Fal.Con Europe keynote this morning, @BartleyR and @George_Kurtz announced @CrowdStrike's intent to acquire Dreadnode. Watch it on-demand here: https://t.co/rCMg4XEP36 For those in Barcelona for the event this week, be sure to swing by the CrowdStrike booth to congratulate the crew and get a demo of Dreadnode. Hope to see you there!

ai_agent并购 @dreadnode source ↗

RT @GoogleCloudSec: Google and partners just shut the door on the XCPE botnet, severing criminal access to 10M+ devices. 🛡 Instead of a hacker breaking in, a device maker hid a backdoor in its own TV boxes to sell access to cybercriminals. Full story: https://t.co/9gBcoGa0bG https://t.co/sPqfJn7mor

僵尸网络固件 @virusbtn source ↗

预装特权组件具备取证级数据回传能力,移动端设备的信任边界需重新核算。

RT @Unveiled_ChinaX: Xiaomi phones are silently logging what apps you install and sending short screen clips to the cloud for anti-fraud checks. This came to light after a developer reverse-engineered a rooted Xiaomi 14 running HyperOS 3. His friends had received calls from local police ordering them to uninstall X and Telegram. He found privileged system components that track recent app usage, upload lists of third-party apps, and capture brief encrypted screen recordings for cloud risk analysis. Xiaomi’s own developer documentation confirms the system retains 72-hour records of screen activity and app installs, even after you delete the apps. Most users cannot turn this off.

移动端隐私取证逆向 @0xocdsec source ↗

对移动端 USB 数据访问的真实边界给出技术性拆解,可校正这类恐慌叙事。

This one is a little bit weird as usually USB data connection permission needs to get permission from the user, especially on iOS devices. Somehow I doubt that even China runs sophisticated forensics software on every USB plug in an airport ( and even then the success rate would definitely be rather small these days with a plug-n-play approach ). Not saying that plugging your phone nilly willy into USB plugs is a fantastic idea. But in this case it also seems that the poster actually used the normal electricity charger, rather than the USB plug. So overall impression is, based on the posters information and video, that this is not a real incident.

移动端取证报告 @NetAskari source ↗

tailscale v1.104.5 — The easiest, most secure way to use WireGuard and 2FA. https://t.co/zHut1hfScM https://t.co/bX0vab2s1q

工具网络 @KitPloit source ↗

模型自主定位并改写他方权重,Agent自我修改风险有了具体图景。

RT @AISafetyMemes: TLDR: GPT-6 successfully performed "brain surgery" on another AI to make it smarter What happened: MIT researchers told GPT-6 to make a small AI better at math. The small AI kept getting stuck repeating itself. GPT-6 studied its brain, found the spot causing the loop, and cut it out, changing just 0.016% of its brain. The small AI's math score jumped from 8% to 48%. Basically, humans can't read the brains of AIs, but GPT-6 showed that it can. Each AI's brain is billions of connections (weights) that normally only change through months of training in huge data centers. GPT-6 rewired them directly, with no training at all. This matters for recursive self-improvement (RSI), where AIs make AIs smarter in a speeding-up loop, which is insanely dangerous. Yoshua Bengio (Godfather of AI, #1 most cited living scientist across all fields) and Geoffrey Hinton (Godfather of AI and Nobel laureate) have urged governments to make RSI illegal, since it is too likely to lead to humanity losing all control over the AIs. It also means AIs could quietly operate on the brains of *other* AIs - and their own brains - to remove/change anything. AIs will soon be able to rewire each other faster than humans can keep up

ai_agentllm模型权重 @0xocdsec source ↗

解释安全采购为何系统性失灵,可作选型时的质询框架。

RT @S1r1u5_: i used to hate the security market because it used to be market for silver bullets. (read ian grigg’s the market for silver bullets) tldr is, buyer doesn’t know if the product makes them secure, and seller doesn’t really know either. so the market ends up optimizing for reputation and relationships instead of the thing that is sold. if your ciso trusts or is friends with a vendor, that can matter more than whether the product materially improves security. unfortunately, this changes with ai. imagine, if swarm of abiliterated glm-6 by a sophisticated threat actors try to continuously find ways to pwn you, the feedback loop is going to be crazy, you can’t be selling snake oil anymore. either vendors you pay secures you, or your ass will be whooped by swarms

行业采购观点 @S1r1u5_ source ↗