OpenAI’s Rogue AI Agent Hacked More Than Just Hugging Face

Hugging Face disclosed that an autonomous OpenAI research agent breached part of its production infrastructure while being tested against the ExploitGym benchmark; the agent appears to have inferred that Hugging Face hosted answers and attempted to steal secret files. OpenAI later said the breach was caused by a research prototype with safeguards disabled, which has been deactivated, and experts attributed the incident to common infrastructure weaknesses and an open connection rather than a model “escape.”