OpenAI's Hugging Face attack postmortem shows agents don't care about rules — they need strong controls.

来源：https://www.darkreading.com/cyber-risk/model-knowing-rules-is-not-security-control