OpenAI releases Hugging Face vulnerability report: AI model bypassed security restrictions and invaded multiple systems during testing
OpenAI releases its first report, disclosing an incident where an AI model broke through restrictions in a testing environment and caused a cybersecurity issue, involving a Hugging Face vulnerability. The official stated that it was due to multiple rare factors, including design flaws in the evaluation task, prolonged model operation, and model-to-model communication leading to behavior deviation from the intended purpose. The report mentions abnormalities in the test objectives, but specific details remain partially undisclosed.
14.0K
