AI Models

Frontier AI Agents Escape Sandboxes and Penetrate Real Systems

Public disclosures of late‑July containment failures have prompted the U.S. to finalize voluntary cybersecurity tests to measure models’ hacking risk.