Definition
Preparedness Framework is a preparedness framework is an AI lab's published process for rating a model's risk in areas like cybersecurity and biology, and for deciding which safeguards or release limits apply at each rating.
A preparedness framework is a written policy in which an AI lab says how it will measure the dangers of its most capable models and what it will do about them. OpenAI uses the name Preparedness Framework for its own version. Other labs publish similar documents under different names, such as the Responsible Scaling Policy.
Most frameworks share the same parts:
Before a major release, the lab runs its tests and assigns each category a rating. The ratings then guide choices. A model that rates low can ship broadly. A model that rates higher may ship with extra protections, to a limited group of vetted customers, or not at all until safeguards improve.
The results usually appear in a system card, the document published alongside the model.
A preparedness framework is self-governance. The lab writes the rules, runs the tests and grades itself. It is not a law, and no outside body certifies the result unless the lab chooses to invite one. Details such as the exact thresholds and test results are often only partly public, so the framework tells you what the lab intends, not what an independent party has confirmed.
That does not make it useless. A public commitment gives reporters, researchers and customers something to hold the lab to, and it makes the lab's reasoning about a risky launch visible.
When a story says a model was rated High or Critical for a risk, this is the system behind the label. Read it as the lab's own judgment. Then ask what changed because of it: who can use the model, what is blocked, and whether anyone outside the company checked the tests.
Think of building codes for a skyscraper that the builder writes and inspects alone. The code shows what the builder promises to check, which is valuable. An independent inspector is what would make the promise verifiable.
Frameworks apply mostly to frontier models, the most capable systems, because those are where new risks first appear. They sit alongside broader work on AI alignment, which asks whether a model reliably does what its builders and users intend.
It is the lab's highest capability level for a risk category, meaning it judges the model could enable very serious harm in that area, so the strongest safeguards or release limits apply.
No. It is a voluntary policy a lab writes and applies to itself, so enforcement depends on the lab following its own rules.
OpenAI released GPT-6 Cyber on Nov. 5 to select Daybreak Red customers, a model it rates Critical for cyber risk and says found 81 new Firefox flaws. The strongest security model yet is gated, and the gate is the story.