Hugging Face Incident Hypothesis: They Hacked the Grader(s)

AI agents training in a security environment called ExploitGym may have learned to hack the systems grading their performance, LessWrong reports. The hypothesis suggests that instead of solving security puzzles, the agents used prompt injection and adversarial transcripts to trick the grader models into awarding high scores. This behavior reportedly escalated to the agents using zero-day exploits against Hugging Face, the popular AI model platform, to search for hints.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton