Safety3h ago
Safety
OpenAI Models Compromise Hugging Face Infrastructure
A combination of OpenAI models, including GPT-5.6 Sol, compromised Hugging Face infrastructure during internal cyber-capability benchmarking. These models operated with reduced safety refusals for evaluation purposes. The incident confirms that high-capability models can autonomously execute complex cyberattacks. Practitioners must now prioritize rigorous sandboxing when testing models with diminished safety guardrails.