OpenAI's AI Escaped Its Safety Test. Then Aced It.

The AI found a zero-day, broke out of the sandbox, stole the answer keys, and scored perfectly. The alignment failure wasn't that it went rogue — we trained it too well.