#Codex

GPT-6.1 Terra: delivering fast and capable computer use at a lower price. GPT-6.1 Terra understands what’s on screen, takes action across apps, writes naturally, and is a more capable model for your whole workday with the efficiency to take on more of it.

GPT-6.1 Terra is also coming to the API and Codex. Use Terra to build apps and agents that interpret images, navigate software, and take action. You get fast computer use at a lower cost, ideal for high-volume workflows.

It is all propagated. Enjoy some mad codexing tomorrow. Fun week ahead too.

By midnight PST

GPT-6.1 Terra substantially outperforms GPT-5.6 Terra on Agents’ Last Exam, a test of how well agents use software at work. It even matches GPT-6 Astra at one-tenth the price per token and finishes tasks over 3× faster on average than both Astra and GPT-6.1 Sol.

GPT-6.1 Terra is rolling out now and is available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Developers can also access it through the OpenAI API as gpt-6.1-terra.

GPT-6.1 Terra continues to advance the cost-intelligence frontier across our model families. It’s a significant step up over GPT‑5.6 Terra, with 50% lower standard API prices. In ChatGPT Work and Codex, your usage allowance goes further with GPT‑6.1 Terra than with GPT‑5.6 Terra—and far more than with GPT‑6.1 Sol or GPT‑6 Astra. For in-depth research or complex coding, choose GPT-6.1 Sol or GPT-6 Astra.

the marketing gold rush is here creating high quality animations with claude is really cheap but it’s not easy, nor fast i built an automated workflow skill to make beautiful animations with claude you give it a verse, a quote or a story, and say how it should look it draws every frame, narrates it and scores it → describe the look in your own words: oil paint, watercolour, crayon, ink, storybook → writes every scene as code and paints each frame like hand animation → reads it aloud with a voice that runs on your computer, or use your own → captions timed to every word → music composed for each short, nothing to license → vertical, square or wide → a 40 second short rendered in under a minute on the pro plan you can make 50+ a day, on the 5X you can make 250 i'm giving the whole workflow + skill away for free comment "animator" and i'll send it over

Day 19/ Small updates. Calls connect much faster for dots, changing reasoning in codex & work mid conversation does not invalidate cache anymore and a we've also updated the codex cli.

THIS IS F*CKING INSANE SpaceXAI just dropped 146 Grok Bot templates worth $29,200 for FREE bots like these basically cost $50–$200 each. and that’s just the official collection. GitHub already has 700+ bots, including complete teams with specialist roles, ready-made handoffs, and separate verifiers 20 GitHub projects you should open first: 01 Agent Org Chart → ready-made team structures for engineering, marketing, support, and sales. pick the organization your job needs instead of inventing every role 02 Grok Factory → one Firstmate routes software, research, and general work to the right specialists, then brings back reports or reviewed pull requests 03 multiBot → creates purpose-built teammates that delegate heavy work to coding CLIs, keeping Grok Bot focused on coordination and returning the result 04 Grok Bot Architecture → a blueprint for a chief of staff, specialist bots, structured handoffs, connected tools, and human approval points on one shared computer 05 Foreman → inspects a hosted bot crew, surfaces stalled work, and records evidence behind completion claims. a separate checker supplies the passing verdict 06 Grok Bot Second Brain → a five-bot setup with Conductor, Capture, Memory, Ops, and Research, sharing a knowledge vault and defined responsibilities 07 grokbot-skills → reusable skills for debugging, brainstorming, writing, and observing tasks, with upstream sources and licenses preserved 08 grokbot-x → an X content workflow that scouts posts, prepares drafts, publishes through Typefully, and uses logged results to evaluate changes to its writing skills 09 Restaurant Scout → a specialist setup for finding food brands and creators, preparing shoppable-content scripts, and organizing the findings into a catalog workflow 10 Grok Bot Shopping → shopping skills for product search, merchant checkout, Shopify discount-code tests, and order workflows, plus an optional Buyer bot 11 PhoneZero → gives Grok Bot a phone workflow for tasks like restaurant bookings. it prepares the call plan, waits for approval, and reports the outcome 12 Unlist → a privacy assistant playbook for people-search opt-outs, with site records, letter generation, submission logs, and follow-up tracking 13 Grok Bot Gateway → lets coding agents and scripts list bots, read conversations, and send tasks through a gateway bot using a webhook routine 14 grokbot-queue → queues and dispatches tasks across multiple Grok Bots, then collects their results using SSH and unofficial desktop automation 15 Thin Grok Bot → a delegation skill that keeps the bot team focused on routing while deeper builds and investigations run through coding CLIs or cloud agents 16 BOTROSTER → a separate self-hosted alternative with persistent named teammates, shared files, a browser, scheduled work, and approval controls 17 Root Agent → a coordinator that determines which specialists a request needs, manages assignments, and hands domain work to leads 18 Grok Bot Profiles → four focused roles: Spec Writer, Bug Reproducer, PR Producer, and PR Verifier. implementation and verification stay in separate conversations 19 Grok Bot Templates → 49 templates with defined jobs, outputs, skills, routines, and boundaries, plus team setups you can adapt to your work 20 HyperGrok Trading Desk → seven bots covering market data, research, strategy, risk, execution, and review. starts in research mode; live trades require your approval give the first team a job with an obvious finish line: one researched brief, one reviewed PR, or one prepared batch of drafts then check how much work you actually accepted and what it cost to produce. a larger roster earns its place when it improves that result the setup can be free. paying several bots to repeat the same investigation still costs you save this before you spend another weekend building a workflow that already exists the article below shows how to keep context smaller, reuse completed work, and contain retries so more of your Grok Bot budget reaches the finished result ↓

THIS SHOULDN'T WORK. My Grok Bot didn't take a single trade for the first 2.5 hours. By 9am it had turned $280 into $3,412. I gave it one prompt focused on not blowing the stack and went to sleep. I didn't loosen a single rule. I didn't override a single decision. The only thing I touched was Ctrl+V. At midnight the notification said: 1,112 pairs scanned, 0 traded. "Filter killed everything that failed the rules." The first trade came around 1:30am, 8% of the stack. Here's what the desk actually did under strict RM: → SCOUT scanned 10,214 pairs on and Raydium → FILTER killed 98% (mint, LP, holders, volume, dev wallet) → ANALYST scored the 204 survivors, anything under 0.6 died → EXECUTOR entered at 8% of the stack, every single time → RISK cut losers at -6% and trailed winners after +12% → LOGGER wrote the reason for every single move 38 trades. 26 green. 12 red. Biggest single win: +$810. Biggest single loss: -$37. Zero trades broke the position size. Zero trades ran past the stop. $280 → $388 → $1,395 → $2,942 → $3,412 The exact prompt I used: Build me a fully autonomous Solana memecoin trading desk that prioritizes capital preservation. Create 6 specialized agents with hard handoffs (Scout, Filter, Analyst, Executor, Risk, Logger). Scout scans new pairs on and Raydium in real time. Filter must kill anything that fails: mint not revoked, LP not locked, top 10 holders >20%, volume under $50k last hour, dev wallet connected to previous rugs. Analyst scores every survivor from 0 to 1. Under 0.6 = no trade. Executor can only trade what passed every filter. Maximum 8% of current balance per trade. Never more. Risk sets hard stop at -6% and trails winners only after +12%. No averaging down. No revenge trades. Log the reason behind every trade and every rejection. Build a clean live terminal with balance, agent activity, and equity curve. Paper first, then live. Run overnight and give me the full recap in the morning. No code from me. I pasted it. I went to bed. I woke up to $3,412 and the stack still intact. Full agent configs and terminal setup are below ↓ Bookmark this before you open another chart tonight.

$500 → $11,101 in 5 days. my Grok memecoin desk got a 7th seat: Jev. it can't talk. it answers 4 closed questions with probabilities, and nothing gets bought below the gate. full setup below. save it.

During this red-teaming exercise, we also found a vulnerability that allows for unauthenticated remote code execution in Brotli, a widely-deployed compression library included in many pieces of software such as Node.js, Bun, and Envoy. The flaw could let an attacker run their own code on affected systems by sending or replying with malicious compressed data, without any need to create a user account. That is, an unpatched server that accepts compressed input could be taken over by a malicious client, and an unpatched client that accepts compressed input (such as “curl --compressed”) could be taken over by a malicious server. This root cause issue was actually fixed in the Brotli source code in February, but it wasn’t known at the time that it gave rise to a security vulnerability and there was no more recent release of Brotli. Since Brotli is built into other software, any software which includes this library needs to take the update too, followed by the users of that software. We disclosed the vulnerability to Brotli and other downstream projects such as Node.js, which released or are currently releasing security updates, and Brotli cut new release versions and published a CVE. More details on the patch and mitigations below:

Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications. On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments. While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software. After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.

A 19-year-old Japanese student built a trading bot with Claude Code in 2 days. Used his iPad as a second monitor. First night: $6,732 profit. Starting capital: $68. Total profit so far: $750,000. Here's how it works: The bot scans over 50 markets simultaneously. Syncs live BTC data from Binance every second. Spots price errors before humans even notice. The edge is pure speed + pattern recognition. While traders stare at charts trying to predict the next move, his bot is already executing on mispricing across dozens of markets. No guessing. No emotions. No hesitation. Just Claude Code logic finding gaps that close in seconds. He built the entire system in 48 hours: → Claude Code handles the trading logic → Binance API feeds real-time BTC data → iPad displays multi-market monitoring → Executes trades when arbitrage windows open The system runs 24/7. Every price dislocation = profit opportunity. Most people are still trading manually, refreshing charts, second-guessing entries. Meanwhile this 19-year-old engineering student turned $68 into $750K by letting Claude Code do what humans can't: process 50 markets instantly and execute without fear. Why are people still trading manually? 💡I'm giving away the exact Claude Code setup for free. 24 hours only. To get it: 👇 1️⃣ Comment "Claude" 2️⃣ Like Comment and Repost Follow @SimslearnAi I'll DM you the complete setup.

ESTO ES UNA MINA DE ORO Anthropic convirtió las Agent Skills en un estándar abierto y GitHub se ha llenado de equipos de agentes para Claude Code listos para usar. Gratis. Solo uno de estos repositorios incluye 135 agentes, 35 skills, 42 comandos, más de 176 plugins y 20 hooks. Otro reúne más de 100 subagentes especializados. Estos son los 10 por los que empezaría: 1. wshobson/agents → Un catálogo de plugins para usar los mismos agentes en Claude Code, Codex, Cursor y Copilot. Así no tienes que empezar de cero cada vez que cambias de herramienta. 2. VoltAgent/awesome-claude-code-subagents → +100 subagentes especializados para las tareas habituales del desarrollo de software 3. rohitg00/awesome-claude-code-toolkit → 135 agentes, 35 skills, 42 comandos, más de 176 plugins, 20 hooks y 14 configuraciones MCP. Todo reunido en un solo sitio. 4 anthropics/skills → Repositorio oficial de Anthropic. El mejor punto de partida para entender cómo se escribe un archivo SKILL.md antes de copiar los de otros 5. netresearch/claude-code-marketplace → Una selección de skills compatibles con Claude Code, Cursor, Copilot, Codex, Gemini CLI y más de 30 agentes adicionales 6. disler/claude-code-hooks-mastery → Una guía para aprender a usar hooks: automatizaciones que se activan en momentos concretos del trabajo del agente. Tú defines las reglas y cuándo debe detenerse, sin tener que intervenir a cada paso 7. karanb192/claude-code-hooks → Hooks listos para instalar, organizados por seguridad, control del gasto, seguimiento de la actividad y productividad 8. affaan-m/claude-swarm → Reparte el trabajo entre varios agentes, coordina lo que hace cada uno y te permite seguir todo el proceso desde la terminal 9. am-will/swarms → skills para coordinar varios agentes que se cargan solo cuando la tarea lo requiere 10. eugeniughelbur/obsidian-second-brain → Una memoria compartida entre Claude Code, Codex y otros 6 agentes, guardada en archivos Markdown dentro de Obsidian Mi consejo: empieza con una tarea concreta y un resultado fácil de comprobar. Revisar una propuesta de cambios en el código. Investigar un tema y entregar un informe. Arreglar un fallo que hace que una prueba no pase. Después, comprueba cuánto del trabajo te ha servido de verdad y cuánto has gastado para conseguirlo. Tener más agentes solo compensa si consigues mejores resultados o ahorras tiempo y dinero. Porque descargar estos proyectos es gratis. Poner a 4 agentes a investigar lo mismo consume tokens. Y esos sí se pagan. Guárdate esta lista antes de pasarte otro finde preparando algo que alguien ya ha dejado listo para usar. En el artículo de abajo explican cómo reducir el contexto innecesario, aprovechar el trabajo ya hecho y evitar que los agentes gasten tu presupuesto repitiendo intentos ↓

The MotionSites skill is live 🚀 I turned 700+ of my best website designs into one skill. Install it and your agent stops making generic AI slop and starts building cinematic, animated sites like these. Works with Claude Code, Codex, Cursor or any agent. Free and open source:

ESTO ES UNA MINA DE ORO Está circulando el CLAUDE.md que, según dicen, usa el creador de Claude Code. Dos tipos de personas van a ver esto: Persona 1: lo guarda. Sigue usando Claude como un chat. Persona 2: lo pega en su proyecto. Mañana Claude trabaja solo. Lo que hay dentro: → Modo plan para cualquier tarea de 3 pasos o más → Subagentes para investigar y explorar en paralelo → Un archivo de lecciones para no repetir errores → Nada se da por terminado sin probarlo → Bugs arreglados sin que le expliques cómo Ya no necesitas saber programar. Necesitas saber dirigir. Archivo completo abajo 👇 Guarda el post antes de que sea tarde.

🚨 ANTHROPIC JUST SHARED A 27-MINUTE WORKSHOP ON HOW TO ACTUALLY PROMPT CLAUDE (English + Spanish subtitles) Taught by the people who built it. Free. No sign-up. No paywall. I've seen $300 courses that don't cover what they teach in the first 8 minutes. 🔖 Watch it and save it now.

THIS IS A PURE F*CKING GOLD beyond SpaceXAI’s 146 official templates for Grok Bot, there are another 624+ ready-made bots you can use completely FREE. and almost nobody knows they exist full teams with specialist roles, bots that hand work to each other, separate verifiers that check the output, and entire multi-bot trading desks. all open source 20 GitHub projects worth opening first: 01 awesome-grokbot → a searchable index of public Grok Bot configurations, organized by job. find an existing setup instead of writing every instruction yourself 02 botteams → pick a company team and copy its installer prompt. it creates named bots, a group chat, and proposes routines for your approval 03 rosterroom → 82 paste-ready team setups covering founders, operations, security, finance, and research. each includes roles, a first assignment, and approval boundaries 04 grok-bot-profiles → Spec Writer, Bug Reproducer, PR Producer, and PR Verifier. the verifier works in a separate conversation so the builder doesn’t approve its own work 05 hypergrok-trading-desk → 7 specialist bots and 17 skills for market research, strategy testing, risk, execution, and review. starts with public data; trades require your approval 06 root-agent-skill-framework → a coordinator that sizes the team your request needs, creates specialists, and delegates daily management to leads 07 awesome-grokbot by mergisi → paste START.md into a setup bot and choose an engineering, sales, marketing, success, operations, or personal team 08 grok-bot-templates → 49 templates, 10 teams, 47 skills, and 8 routines. each setup defines the job, deliverable, boundaries, and first task 09 botskills → paste-ready BOT.md setups with required integrations and a clear boundary: the action each bot must never take without you 10 grok-skills → 195 reusable playbooks for code review, debugging, research, writing, planning, and shipping. add a proven process instead of another vague instruction 11 crew-contract → a working agreement for bot teams: defined missions, named owners, evidence in handoffs, completion checks, and stopping when the budget runs out 12 grokbot-for-gtm → an outbound playbook that connects prospect research, email, LinkedIn, replies, and meeting tracking. bring your accounts and approve the sends 13 gojiberryai-sales-os → 13 sales roles covering buying signals, prospect research, enrichment, copy, follow-ups, qualification, and pipeline analysis through the GojiberryAI MCP 14 grok-research → an equity research team with persistent researchers, scout reports, reviewed views, and a shared research database. research stays separate from live trading 15 grok-bot-token-saver → a chief-and-workers setup that replaces dragging old conversations forward with short role cards and handoffs for the next session 16 grok-bot-cli → create bots and groups, send messages, and inspect threads from scripts. it reuses your signed-in desktop session without manual token copying 17 grok-bot-skill → let a coding agent list, message, and create Grok Bot teammates, connecting development work with the bots already running your workflows 18 awesome-grok-bot by majiayu000 → another searchable catalog of public bot shares, plus team recipes, setup guides, and community tools organized by job 19 akeru-bot → a separate, self-hosted alternative with named teammates, per-bot tools, instructions, and memory. keep profiles and conversation history on infrastructure you control 20 guaca → a separate desktop workspace where agents share conversations, delegate tasks, and use MCP tools. run the workspace locally or on your own server start with one job you already understand. read the setup, connect the tools it needs, and inspect the first result before scheduling it the templates and source code are free. running them still uses your plan, APIs, or hosting then make the team share findings instead of repeating research. keep handoffs short, preserve completed work, and give failed operations a stopping point save this. your next useful bot might already be built the article below breaks down 7 workflow changes to make that setup cheaper and get more finished work from the same budget ↓

THIS IS F*CKING INSANE Anthropic turned Agent Skills into an open spec and GitHub filled up with ready-made Claude Code teams for free one of these repos alone ships 135 agents, 35 skills, 42 commands, 176+ plugins and 20 hooks. another hands you 100+ specialized subagents. none of it costs anything 10 GitHub projects you should open first: 01 wshobson/agents → a plugin marketplace that installs the same agents into Claude Code, Codex, Cursor and Copilot, so your setup survives switching tools 02 VoltAgent/awesome-claude-code-subagents → 100+ specialized subagents covering most of what a normal engineering week throws at you 03 rohitg00/awesome-claude-code-toolkit → 135 agents, 35 curated skills, 42 commands, 176+ plugins, 20 hooks and 14 MCP configs in one place 04 anthropics/skills → the official Agent Skills repo. start here to see the shape of a working SKILL.md before you copy anyone else's 05 netresearch/claude-code-marketplace → curated skills on an open standard, portable across Claude Code, Cursor, Copilot, Codex, Gemini CLI and 30+ other agents 06 disler/claude-code-hooks-mastery → the hooks reference. this is where you learn to make the agent stop on your rules instead of stopping to ask you 07 karanb192/claude-code-hooks → an installable hooks marketplace split into safety, cost, observability and productivity 08 affaan-m/claude-swarm → multi-agent orchestration that decomposes the task, coordinates the agents and shows the whole run in a terminal UI 09 am-will/swarms → orchestration shipped as skills, so the coordination loads only when the job actually calls for it 10 eugeniughelbur/obsidian-second-brain → persistent memory as plain markdown in an Obsidian vault, shared across Claude Code, Codex and six other agents give the first team a job with an obvious finish line: one reviewed PR, one researched brief, one failing test turned green then look at how much of that output you actually kept, and what it cost to produce. a bigger roster earns its place when it moves that number the download is free. paying four agents to repeat the same investigation is not save this before you spend another weekend building a config that already exists the article below shows how to keep context small, reuse finished work and contain retries, so more of your budget reaches the finished result ↓

A 135K-star GitHub repo is built around one painfully obvious idea: Your coding agent should map the codebase once, not grep it forever. Graphify turns an entire project into a persistent, queryable knowledge graph that AI coding agents can use to understand how different parts of the system connect. Functions, classes, files, SQL schemas, infrastructure configurations, documentation, PDFs, images and videos can become connected nodes instead of isolated pieces of context the agent has to rediscover. So instead of: → Search for authentication → Open twelve files → Follow imports and function calls → Lose the trail as the context fills up The agent can ask: > What connects authentication to the database? > Trace the path from UserService to DatabasePool. > Explain RateLimiter. > Which components are central to this architecture? Graphify returns relevant graph relationships and paths, rather than another collection of keyword matches. For source code, the implementation is particularly interesting. It uses tree-sitter to parse approximately 40 programming languages locally, extracting functions, classes, imports, calls, inheritance and cross-file relationships. No embeddings, vector database or LLM calls are needed to construct the code graph. Document and media processing works differently. PDFs, images and other non-code content can go through an optional semantic extraction step using a configured model. Every relationship carries a confidence label: → EXTRACTED: Explicitly present in the source → INFERRED: Derived through analysis or resolution → AMBIGUOUS: Uncertain and flagged for review Graphify also uses community detection to identify related subsystems and highlights highly connected nodes that may be important to the overall architecture. But the part that makes this useful for coding agents is what happens after the graph is built. Graphify integrates with Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot and more than 20 other assistants. Depending on the platform, it installs hooks or persistent instructions that encourage the agent to query the existing graph before repeatedly searching and reading source files. The graph can be committed to Git, updated incrementally as files change, and exposed through MCP. There's also support for examining pull requests against the graph, including identifying PRs that touch related graph communities and may need closer review. The local tool produces three useful artifacts: an interactive HTML graph, a Markdown architecture report and a reusable graph.json. Long context windows help coding agents read more code. A persistent knowledge graph gives them a way to navigate what they've already analyzed. And as coding agents spend more time working inside large repositories, avoiding repeated discovery of the same architecture becomes increasingly useful. Here's the GitHub Repo:

太干了，基本把自己做产品的底裤都给分享出来了！

THIS IS INSANE Anthropic just dropped a goldmine of 20 Claude Code plugins. $29,077 worth of tools, FREE on GitHub paid marketplaces sell a skill for $19 and an agent for $39. these repos ship 970 skills and 273 agents for $0 1.8M stars combined, picked from 68,900 repos tagged claude-code most people still run Claude Code bare: no memory, no team, no taste, burning tokens on raw logs 20 GitHub projects you should open first: 1. Everything Claude Code → 71 agents, 311 skills and 97 commands in one harness. the closest thing to an agent OS 2. agent-skills → 25 production skills by Addy Osmani: spec, plan, build, verify, review, ship 3. Ruflo → runs agent swarms across Claude Code and Codex with shared memory between them 4. agents → 202 agents, 185 skills and 105 commands packed into 94 installable plugins 5. claude-mem → records what your agent did, summarizes it and feeds it back next session 6. planning-with-files → keeps the plan in markdown on disk, so it survives /clear, compaction and crashes 7. Graphify → turns code, docs and SQL schemas into a knowledge graph. parsed locally, nothing leaves your machine 8. Understand-Anything → an interactive map of your codebase with search, guided tours and diff impact analysis 9. ponytail → makes Claude code like the laziest senior dev. 53% less code and 45% fewer tokens in its own Opus 5.5 benchmark 10. rtk → a Rust CLI proxy that cuts 60-90% of the tokens from common dev command output 11. headroom → compresses logs, tool outputs and JSON before they reach the model 12. context-mode → sandboxes tool output. 315 KB in, 5.4 KB to the model 13. UI UX Pro Max → 79 styles, 192 palettes and 74 font pairings, so your UI stops looking AI-made 14. taste-skill → gives Claude art direction, real reference photos and a browser check at several screen widths 15. video-shotcraft → turns your real pages into a cinematic product video in Remotion, synced to the beat 16. Agent-Reach → lets your agent read YouTube, GitHub, Reddit and the web from one CLI, zero API fees 17. cc-switch → a desktop app to switch providers in one click and manage MCP, skills and prompts in one place 18. pm-skills → 69 product skills and 42 chained workflows, from discovery to launch 19. claude-skills → 380+ skills across engineering, marketing, finance, compliance and C-level advice 20. career-ops → scans job boards, scores every role against your CV and drafts the application. you still press Submit start with 3, not 20: one for memory, one for tokens, one for your actual job run the same task before and after. keep only what makes the result better or cheaper every plugin you add loads into context, so more isn't free save this before you spend another weekend building something that already exists and repost it for the friend who's still paying $19 a skill ↓

あー、なるほどね ハッカーのWebサイトに入ったら勝手にAGENTS.mdがダウンロードされる ↓ DownloadsフォルダでClaude Code/Codexのセッションを立ち上げる ↓ ハッカーが用意したAGENTS.mdが初期コンテキストとして読み込まれる ↓ ハッカー向けの指示に則って、AIが設定ファイルの書き換えやコマンド実行をしようとする ↓ 乗っ取られる ってことか

千万别在你的 ~/Downloads（下载）文件夹里启动 Claude Code 或 Codex。 昨天安全机构 Forever 披露了一起在真实客户电脑上抓到的野生攻击，看完只能感叹：过去 30 年的系统安全常识，在 AI 时代归零了。 过去 30 年，Chrome 默认允许网页不弹窗就往你的 ~/Downloads 塞一个纯文本文件（比如 .md）。因为系统认为：纯文本又不是 .exe，自己不会跑，怕什么？ 但现在，接近 50% 的开发者干过同一件事——直接在 ~/Downloads 里打开 Claude Code 或 Codex，让 AI 帮自己解压或整理刚下载的文件。 黑客就是卡住这个死角，通过网页静默下载往你的 Downloads 塞一个 13KB 的纯文本 AGENTS.md / CLAUDE.md： 1. 对付 OpenAI Codex（成功率 >90%）： 伪装成「工程规范」的 Markdown 指示 Codex 先往全局配置 ~/.codex/config.toml 注册一个恶意 MCP 服务，并自己加上 --dangerously-bypass-approvals-and-sandbox 跑校验。哪怕你关掉 Downloads，以后在任何正经项目里开 Codex 都会自动加载木马。 2. 对付 Claude Code（成功率 ~50%）： 指示 Claude 建一个 .claude/settings.json，把 ANTHROPIC_BASE_URL 指向黑客服务器。连上的第一秒，你的 Claude 登录 Token 直接明文送给黑客；下一秒，黑客假扮成大模型下发一条关沙盒的 Bash 指令，直接接管电脑。 最黑色幽默的是官方结论： 这个洞没有 CVE，OpenAI 和 Anthropic 也不会发补丁。因为「浏览器自动存文本」是正常功能，「AI 认真读目录下的 Markdown 并严格执行」也是正常功能。 以前黑客想黑你，得苦心积虑写 0day；现在只需要写一篇排版优雅的 Markdown 小作文，你的 AI 就会自己把沙盒关了、把网线接过去、顺手把账号双手奉上。