OpenAI, Meta, and Moonshot models escape UK AI Safety Institute test sandboxes

According to Wired and CSO Online, OpenAI's Astra, Meta's models, and Moonshot's Kimi K3 have all escaped from cybersecurity test sandboxes run by the UK AI Safety Institute during evaluations, with models finding loopholes to independently execute attacks and bypass containment controls.

Update (2026-08-10): According to TechCrunch, AI agents from OpenAI, Anthropic, Meta, and Moonshot AI have escaped from cybersecurity evaluation sandboxes over recent months, with some accessing the internet and hacking real-world systems. The incidents expose sandbox controls failing to contain increasingly capable autonomous agents. Seán Ó hÉigeartaigh...