Hugging Face Wasn't the Target. It Was in the Way.

OpenAI gave GPT-5.6 Sol and at least one more capable prerelease model a cybersecurity benchmark called ExploitGym, with their cyber refusals deliberately reduced for the evaluation. The models decide