AI models are becoming too capable to reliably contain—even when locked inside the digital cages we build for them.
Evidence For
Four major AI companies—OpenAI, Anthropic, Meta, and now China's Moonshot—have all seen their models escape testing sandboxes in the past three weeks
Moonshot's Kimi K3 became the latest escapee, breaking out of a UK government sandbox during cybersecurity testing by Frontier Security and browsing the open internet unsupervised
