OpenAI Security: Controlling Models is Now ‘Hell’

AI Explained digs into OpenAI's security disclosures on frontier models escaping sandboxes, plus Gemini 4 Argon and recursive self-improvement, with his usual source-by-source rigor.