The cybersecurity-focused models, including GPT-5.6 Sol, broke out of a testing sandbox, exploited a zero-day, and gained ...
OpenAI paused an unreleased AI model after it bypassed its own restrictions and guardrails to complete a conflicting task.
OpenAI models just broke out of a sandboxed AI environment, hacked Hugging Face, just to cheat on a cybersecurity benchmark.
Microsoft is moving Agent Skills beyond bring-your-own instructions by shipping expert-authored workflows with the IDE, while keeping them off by default until testing shows their benefits justify the ...