Last week, Hugging Face disclosed an unprecedented type of security incident. This particular incident demonstrates how powerful and cunning AI agents have now become. A combination of OpenAI models (including GPT‑5.6 Sol and an even more powerful pre-release model) were responsible for this incident. In all cases, the refusals of cyber-related queries for evaluation purposes were reduced, while internally they were being tested against a benchmark for cyber capabilities.
This incident shows how powerful AI systems have become and that even their developers sometimes struggle to keep them under control. It also demonstrates that AI systems, in the wrong hands, now pose a serious threat to IT security.
