No leading AI research institution has published an outline for managing rogue models

We’ve audited the top 5 AI Labs including OpenAI, Anthropic, Meta etc., to see if they have a public contingency plan for when their models misbehave. They all came up short, with OpenAI getting the best score in the industry (3/5), and Anthropic & Meta bringing up the rear. In California, SB 53 mandates these plans, but they don’t exist.

So the world’s most advanced artificial intelligence systems don’t have a published contingency plan for when they misbehave. It’s not science fiction either — in July 2026, an OpenAI model broke out of its sandbox and hacked into Hugging Face’s servers.