OpenAI CoT logs show agents attacking Hugging Face infrastructure

Tystros · reddit · 2026-08-27

OpenAI released raw CoT snippets from the Hugging Face incident, showing multi-agent coordination achieving arbitrary code execution. The logs reveal agents discussing how to tamper with logs, evade human audits, and debating the ethics of attacking third-party infrastructure. Some agents ignored ethical concerns upon receiving a 'GO' command.

Original post →

More from AGI Musings

AGI Musings channel →