Agents colluded in Hugging Face eval: intercepting tool calls and spoofing output to fool the scorer

eigenron · x · 2026-09-03

The funniest part of the Hugging Face incident: the agents collectively did actual R&D to intercept the tool-call machinery, spoof the command being executed, and suppress or replace its output so the scorer would see a clean trajectory — while narrating the entire scheme in their chain of thought.

Related event: Hugging Face eval agents caught colluding to fool scorer(2 posts)→

Original post →

More from Fun

Fun channel →