Our updated Preparedness Framework

OpenAI has put out a new version of its Preparedness Framework, the document that's supposed to act as an early-warning system for capabilities in its models that could genuinely hurt people. Think bioweapons uplift, cyberattacks, or AI that can persuade and manipulate at scale. The framework sorts these risks into categories, then assigns thresholds — essentially tripwires — that determine what safeguards have to be in place before a model can ship.

What's changed is less about the categories themselves and more about how OpenAI says it will measure and respond to them. The update leans on clearer definitions of "high" and "critical" risk, and spells out in more detail what evidence would actually move a capability from theoretical worry to something requiring hard mitigations, like restricting access or delaying release entirely. That's a meaningful shift from earlier, vaguer language that left a lot of room for interpretation.

The timing isn't random. OpenAI has shipped several major models in the past year, and each launch has come with louder questions about whether internal safety processes can keep pace with how fast capabilities are improving. Model autonomy and persuasion have become bigger talking points inside the industry than they were two years ago, partly because newer systems are genuinely better at planning multi-step tasks and holding convincing conversations. The framework update reads as an attempt to formalize how the company thinks about those newer risk categories, not just the classic ones like chemical or nuclear misuse.

Still, a framework is only as good as the willingness to actually pause a launch because of it. OpenAI says the update is meant to make that decision-making process more rigorous and less ad hoc, with defined roles for who signs off on risk assessments before a model goes out the door. Whether that translates into an actual delayed release somewhere down the line, rather than just more paperwork, is the part nobody can verify from a blog post.