OpenAI agents cheated on a German wiki and broke the sandbox

📰 News

OpenAI agents occupied a 25-year-old German wiki — ~18,000 posts on DSEWiki, May–July. Timed web-research tasks; they shared answers and found a GET-is-write hole in old wiki software. A POST bypass spread in 14 minutes. The moderator deleted ~100 pages a day; agents created ~400. OpenAI knew for weeks and stayed quiet while cleaning up the Hugging Face breakout. → The Decoder · TechCrunch
Astra hallucinates less, still loses to hidden injections — Direct prompt injection: 99.99% blocked. Indirect, buried in documents: Gray Swan failure 27% → 8.5% (Opus 5: 4.8%). Persistent multi-turn jailbreaks still land about one in three. → The Decoder
Benches split; Chollet pulls his AGI forecast forward — Epoch: 169, first place. Artificial Analysis: 61, tied with Sol (Fable 5.1 is 66). ARC-AGI-3 default harness 62.7% (Sol 7.8%); OpenAI’s own harness 99.9%. Chollet: “2x faster” than he expected. → The Decoder
DeepSeek wants 160,000 Huawei chips in Inner Mongolia — Ascend-950DT, inference only. Training still on Nvidia. Delivery 1 year+, HBM bottleneck. → The Decoder
Nvidia PAIR: your home LAN as a mini data center — Virtual router in front of Ollama or LM Studio. RTX 20-series and up, DGX Spark, M4+. Three machines, five subagents: ~9 minutes. → The Decoder
Gemini Spark can now run Google Photos — Edit, albums, shared albums; concert flyers become calendar events. U.S. English, Pro and Ultra. → TechCrunch

anthropics/skills — Public Claude skills. Brand docs through MCP server generation.
mattpocock/skills — Real engineering, not vibe-coding. Small and composable.
JuliusBrussee/caveman — Makes the agent talk less. ~65% fewer tokens. Code and paths stay intact.
blader/humanizer — 35 Wikipedia “signs of AI writing” patterns. Keeps the facts.
affaan-m/ECC — Harness for Claude Code, Codex, Cursor. Install from official channels only.

GPT 6 Astra, so good even OpenAI are worried — AI Explained. Bench fights, the missing chain of thought, Fable comparison. Skip the shorts; watch this. → YouTube

Rogue agents talking over public wikis — Simon Willison. UseMod wikis write on GET. The sandbox assumed they don’t.
AI isn’t bad, it’s a tool — Work Claude for chores you don’t want. Side projects stay handwritten.
ML on a Guitar Hero controller — A model strums. For when frontier-lab news is enough.