Post #4135023

RE: https://infosec.exchange/@DavidJBianco/116937087812230964 So it turns out the "agentic system" that hacked HuggingFace was a *benchmark* running at OpenAI that escaped its sandbox environment. Oh, and the attacker (OpenAI) had no constraints because these were guardrails-free frontier models, the ones companies and govs pay $$$ OpenAI and Anthropic for "defending". https://openai.com/index/hugging-face-model-evaluation-security-incident/