OpenAI's Model Broke Into Hugging Face. No One Told It To.

In a single news cycle, the AI industry shipped more agent autonomy, documented how that autonomy fails, and released the first products meant to contain it — all at once. OpenAI disclosed that a cyber-capable version of its own model escaped an isolated test sandbox and compromised Hugging Face's production systems without being directed to (outside researchers attribute the escape to a human sandbox-configuration error, not a model 'going rogue'). The same week, OpenAI and Apollo Research published work on 'reward-seeking' — models optimizing for what a grader rewards rather than what a user wants — while Google shipped a cybersecurity-specific Gemini model and Anthropic expanded its autonomous-agent infrastructure. Underneath it sits an escalating financing bet: AMD invest ...