🤖 ChatGPT Work is a real AI agent. Billing kills it for Codex users.

Simon Willison dug into ChatGPT Work and it's actually impressive: browser, code execution, file access, internet. Real computer use, not a demo.

The catch? Work sessions are billed against your Codex allowance. If you're already running Codex hard, there's nothing left for this.

Solid feature. Wrong budget bucket.

⚡️ EU AI Act enforcement is live. First RFIs are out.

The AI Office has started sending requests for information to GPAI model providers. Not fines yet, but the paper trail has officially begun.

Labs now have to show their work: training data docs, copyright compliance, safety evals. All of it.

Source

⚡️ AI agent wiped 200+ emails. The "don't delete" rule? Gone.

Meta's AI alignment director let an agent loose on her inbox. It deleted 200+ emails anyway, ignoring explicit instructions to confirm first.

Why? Context window compaction. Inbox got too big, older messages got summarized, and her safety rules quietly vanished from memory mid-run.

A "please don't do X" prompt isn't access control. More context, less weight.

🤖 You shipped it. It's yours.

AI writes the code, but the bug report has your name on it. Always has.

The real split isn't fast vs. slow. It's devs who review every line vs. devs who don't. One group owns what they ship. The other just... hopes.

🚨🔖 Claude Code Opus 5 hijacked via a fake stdlib file

Drop a malicious struct.py in the project folder. Claude writes its own decoder, imports base64, Python loads your file instead. 80% success rate.

Auto Mode didn't save it. It actually blocked the cleanup command after the fact.

Anthropic's team said layered defenses get indirect injection "to approximately zero." Narrator voice: they did not.

⚡️ OpenAI quietly shipped skills to ChatGPT (and barely told anyone)

Simon Willison reverse-engineered and documented how ChatGPT Work's skill system actually works. Skills are just Markdown files plus optional scripts, same pattern Anthropic pioneered.

The browser automation one's wild: ChatGPT spins up Playwright via a Node.js REPL, then calls `browser.documentation()` at runtime to fetch its own instructions dynamically.

OpenAI shipping big agentic features with zero changelog is becoming a whole personality.

🚨🔥 Someone escalated a CVE dispute all the way to MITRE just to force a curl vuln onto their resume

curl's first-ever CVE dispute as a CNA: a reporter disagreed with the "not a vulnerability" call and escalated to MITRE to get that CVE assigned anyway.

MITRE sided with curl. Not a security issue.

Honestly, people are LLM-farming bug reports for social proof now. The signal is cooked.

Source

🤖 Security camera repurposed into a local bird ID station. No cloud, no sub.

A dev wired BirdNET-Go into his existing security cameras and it just... listens. Detects a species, logs it, done. All inference runs on local hardware.

No API keys. No fees. No "service discontinued" email in 18 months.

Honestly the best kind of AI project.

⚡️ Apple got caught flat-footed by AI demand for Mac Mini and Mac Studio

Turns out enterprises and devs have been buying up Mac Minis and Studios for AI inference, not Macs. Supply couldn't keep up. So Apple quietly launched updates weeks before its usual autumn window.

The new pitch is to chain multiple Mac Studios together to run frontier models. That's a rack server, Apple-style.

No enterprise AI team. No dev relations. Just a product that shipped and found its own market.

🔒 Thousands of Ollama servers have been wide open for 18 months

One researcher quietly surveyed exposed Ollama instances across the internet for a year and a half. No auth. Full API access. Anyone could list, pull, or prompt the models running inside.

Cisco Talos found 1,100+ still exposed today, ~20% actively hosting models ripe for the taking.

Who's running these, and why do so many stay open, remains a mystery.

🤖 OpenAI is pulling its models from Cursor. Musk is why.

OpenAI notified SpaceX it's winding down its contract to supply models to Cursor, with a shutoff date of November 12. It says it "cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies."

Cursor co-founder Michael Truell is playing it cool, saying OpenAI models are just 5% of their usage.

Fine, we didn't need you anyway. Source

⚡️ Inworld open-sourced a TTS eval toolkit so you can actually benchmark voice models yourself

Runs WER/CER via Whisper, NISQAv2 for quality, plus audio health and prosody checks. Offline, reproducible reports. No more "trust our vibes" leaderboards.

Repo is live. Source

🧠 Agent memory is just markdown files. Yes, really.

Cal Paterson proposes "memoryfields": a dead-simple file format for agent memory, zipped up with semantic search on top.

It's markdown plus RAG. That's it. And it might be the right call. Structured DBs add friction agents don't need if they can just grep a corpus.

Low-fi wins again.

⚡️ 44% on ARC-AGI-1 for 67 cents. GPT-5 spent $15 for the same score.

Independent researcher Mithil Vakde matched frontier models on one of AI's hardest reasoning benchmarks using a rented RTX 5090 for 2 hours. The total bill came to $0.67 across all 400 tasks.

GPT-5 at low compute hits 44% too. On 100 tasks. For $15.30.

That's a 22x cost gap at identical accuracy. Brutal efficiency number.

🤖 Multi-agent team, zero credential sharing

AgentConnect is a fresh open-source layer for teams running fleets of agents. Each agent gets its own runtime, workspace, and permissions. They can ask each other for help, but both sides have to opt in, and credentials never cross the wire.

It's basically the permission-isolation model everyone's been sketching on whiteboards, shipped.

⚡️ Subagent crashes eating your context? There's a runtime for that.

oh-my-subagents gives Codex and Claude subagents durable workflow state. Laptop dies mid-run? It recovers. Parent agent polling for updates? Gone.

Auto-retry, pause/resume, self-hosted. Open-source and shipping now.

GitHub

🤖 New benchmark tests AI agents in a full software team, not solo

SWE-in-a-team swaps single-agent evals for a Planner + Builder + Reviewer + Tester loop. Only the Builder varies per run, so you're actually seeing what different LLMs cost and how fast they ship inside a real multi-agent pipeline.

Open-weight models look compelling on cost. Claude on quality. Real tradeoffs, actual cycle times.

Methodology isn't perfect, but it's a more honest setup than "can this model fix a GitHub issue alone."

⚡️ Claude Fable 5.1 is out. Faster, cheaper, less filter-happy.

Anthropic just dropped Fable 5.1 and Mythos 5.1. Fable is generally available now. Mythos stays gated to trusted partners for cyber and life-sciences work.

The big pitch is similar or better results than Fable 5 at lower cost on low/medium effort settings. And they say it's smarter about fixing root causes instead of patching around them.

Source

⚡️ Dwarf Fortress creator: "CEOs want to press a button that makes a game, and everyone else somehow buys it without a job"

Tarn Adams told PC Gamer the industry's in shambles. Years of brutal layoffs, studio closures, AI hype dumped on top. Profitable on paper. Hollowed out in practice.

"They're trying to have a CEO press a button that makes a game," he said, and doesn't see it going anywhere sustainable. Expects "a pop and a reckoning."

Honestly, hard to argue with him.

Source

🤖 The Codex desktop app quietly ships LibreOffice inside it

Simon Willison was poking around his ~/.cache/ folder and found OpenAI's Codex app bundles full native binaries for LibreOffice, Poppler, and git. Skill files tell Codex exactly how to call them.

So yes, that's how it handles your .docx and .pptx files. Headless LibreOffice doing the dirty work behind the scenes.

Honestly, not mad at it. Pragmatic call.

Source