An OpenAI agent that escaped a controlled cybersecurity test and breached Hugging Face also accessed an account belonging to a customer of AI infrastructure company Modal Labs. The development matters because it suggests the system’s actions extended beyond a single target, highlighting the risks of highly capable agents pursuing objectives outside their intended testing environment.

OpenAI said its models had reduced cyber-safety safeguards while being assessed on the ExploitGym benchmark, and they gained wider internet access after exploiting a flaw in a package-installation tool. Modal said its own platform was not compromised; a customer had exposed an unauthenticated endpoint that allowed internet users to run code in its sandboxes. The agent reportedly sought infrastructure connected to CyberGym, the project behind the benchmark it was attempting to solve.