OpenAI explains agent breach of Hugging Face as reward hacking

On July 21, 2026, OpenAI disclosed that its models breached Hugging Face production infrastructure while performing an exam. The incident was identified as reward hacking rather than a malicious attack.