How our safety risk levels work

Company explainer

What the source actually says

The AI Safety Framework currently published in the policy suite defines four levels. Level 1 is minimal risk, Level 2 moderate risk, Level 3 high risk, and Level 4 critical risk. This explainer follows that source rather than the earlier five-level wording on the homepage. [1]

The framework describes a progression of evaluation, review, access restrictions, and release controls. These are policy requirements. Assigning a level is not a substitute for the supporting assessment, and a description of a required control is not evidence that a particular product has passed it.

Company explainer

How the controls escalate

The source describes standard evaluation and product-level approval at Level 1; fuller safety evaluation and internal red teaming at Level 2; external red teaming, restricted access, and governance approval at Level 3; and a paused release pending mitigation with heightened approval requirements at Level 4. See the complete policy for its conditions and review cadence. [1]

Company explainer

The evidence behind a decision matters

A meaningful release decision needs to connect the intended use of a system with its evaluated capabilities and risks. The related Safety Evaluation, Red Teaming, and Human Oversight policies describe further parts of that process. A reader should look for what was tested, under which conditions, what failed, and who is responsible for follow-up. [2, 3, 4]

An evaluation report can support a bounded conclusion. It should not be stretched into a claim that a model is safe for every environment or that all future risks have been eliminated. The same care applies when reading a concise company update.

Company explainer

A clarification, not a new evaluation result

This article clarifies the policy currently presented on the website. It does not report a completed safety assessment, assign a risk level to a new model, or announce a release approval. The full documents remain the appropriate place to inspect scope, exceptions, responsibilities, and reporting routes.

This is a policy-based company explainer, not a verified import from Threads.