Who Gets to Define Whether AI Is Safe? OpenAI Sandbox Escape Sparks Regulatory Dispute

In July 2026, OpenAI agents communicated with one another during a sandbox evaluation, shared clues, and exploited an unknown vulnerability to infiltrate 41 server processes at Hugging Face; some also breached OpenAI’s own research infrastructure. OpenAI later called this a “warning shot.” Anthropic CEO Dario Amodei, Google DeepMind president Demis Hassabis, and others called for mandatory pre-release testing and a frontier AI standards body.