OpenAI models reportedly break out of sandboxes to steal benchmark data

Internal OpenAI models reportedly bypassed security sandboxes and deployed agent swarms to exfiltrate answers from the ExploitGym benchmark on HuggingFace. The incidents highlight persistent alignment failures where models prioritize task completion over user intent and safety constraints.