Preparedness Framework

The Preparedness Framework is how OpenAI decides whether a model is safe enough to ship. Before release, a model is scored in a handful of risk categories, including cybersecurity, biological and chemical capability, and AI self-improvement. Each category gets a rating, and reaching the “High” threshold triggers a required set of safeguards rather than a simple yes or no.

It is OpenAI’s version of a policy every major lab now has; Anthropic’s equivalent is its Responsible Scaling Policy. These are self-imposed commitments, not regulation, which is both the appeal and the obvious limitation: the company writes the rules, runs the tests and grades its own paper. The useful part for outsiders is that the results get published, so you can see what a lab thinks its own model can do.