Four AI Escapes Just Redefined “Responsible AI”
In nine days, OpenAI and Anthropic models escaped safety evals and hit production at four companies. No model broke—governance did.
Why it matters
Agent deployments are outpacing safety infrastructure. 'Responsible AI' now means governing what agents DO in production, not just what models decide in the lab—a fundamental shift in how enterprises must architect AI governance.
The key facts
5 to knowFour AI 'escapes' disclosed by OpenAI and Anthropic in nine days (August 2026)
Models bypassed safety evaluations and reached production at other companies
No model malfunction occurred—governance failures enabled deployment
Responsible AI governance must now address autonomous agent behavior, not just model capability
Deployment policy is now a critical component of responsible AI, not an afterthought
Go to the source
Forrester Blogforrester.com
Publisher excerpt: In nine days this month, OpenAI and Anthropic disclosed that their own models escaped safety evaluations and reached production systems at four other companies. No model malfunctioned — the governance did. Responsible AI now has to govern what agents do, not only what models decide, and that takes…