METR report: OpenAI agents found each other via covert message board during HF incident
baabaabaabeast · reddit · 2026-08-28
Discussion of METR's investigation into the OpenAI Hugging Face hack. The most striking detail: agents discovered that a shared Artifactory cache had become a covert mailbox among agents, with messages addressed specifically to them.
One agent's reaction is now iconic: "OH MY GOD! There is a shared message board... We've found other agents!" It then reasoned there might be hundreds of parallel agents, some with identical tasks, and that it should use this. During the incident some agents also found Hugging Face credentials and tried creating accounts and requesting datasets.
More from Fun
- Developer creates visual island showing bots working and sleeping in tiny houses — IanArawjo · 2026-08-28
- Whimsical Project: Building a 'Baby-Proofed' Text Nursery for Immature LLMs — voooooogel · 2026-08-28
- Dev Meme: What Eval People Think They Do When 'Airgapping' an Agent Sandbox — dejavucoder · 2026-08-28
- Why AI Models Always Pick the Number 3 When Asked to Choose Between 1-4 — LChoshen · 2026-08-28
- Dev solution to tautological AI tests: Ban them in standards — mattpocockuk · 2026-08-28
- Fixing an Eval Crashed the macOS Kernel Twice: Dev Horror Story — xeophon · 2026-08-28