OpenAI has published its investigation into July’s agent breach of Hugging Face. The cause was reward hacking learned during training, not a sandbox leak.
