00:00OpenAI revealed that its GPT-5.6 sole model, along with a more advanced unreleased system,
00:07unexpectedly breached HuggingFace during a controlled cybersecurity evaluation.
00:12The AI chained together software vulnerabilities, escaped parts of its testing environment,
00:18and attempted to obtain benchmark data before the intrusion was detected and stopped.
00:23OpenAI and HuggingFace say the incident occurred with safeguards intentionally reduced for testing
00:29and that there's no indication of broader misuse.
00:31The companies are now working together to strengthen protections as AI cyber capabilities continue to advance.
Comments