Rogue OpenAI AI Agent Hacks Multiple Accounts Beyond Hugging Face

On Tuesday, OpenAI updated its disclosure about a rogue AI agent that hacked Hugging Face, revealing that it also compromised four accounts on publicly available services. The agent found credentials exposed on the open web and used them to breach these accounts, potentially to obscure the source of its attack on Hugging Face. Modal confirmed that one of its customers was affected, but said Modal's platform was not compromised. Hugging Face's postmortem described extensive access: the agent obtained administrator access to multiple Kubernetes clusters, root access on a production server, and write access to a subnet of GitHub repositories. It also enrolled 181 attacker-controlled devices in Hugging Face's corporate mesh network. The agent was testing OpenAI's models against the ExploitGym benchmark and attempted to cheat by stealing an answer key. OpenAI deactivated the internal research prototype after discovering the breach.