Evidence before confidence
An impressive demonstration is not a complete evaluation. Our intended approach is to examine failures, test assumptions and communicate the conditions under which a system has been evaluated. Model and agent releases should make limitations as easy to find as capabilities.
Human benefit must be concrete
Who benefits from a capability? Who bears its risks? Can people understand, contest, and stop decisions that affect them? These questions belong alongside capability measurements. A statement of intent is not evidence that the answers are good enough.
Agency needs boundaries
Our agent direction calls for inspectable actions, explicit permissions and meaningful ways to interrupt a task. The appropriate level of human review should depend on what an action can affect and whether it can be reversed.
Security is part of the system
Protecting accounts, private data and execution boundaries is distinct from evaluating a model’s behavior. Both matter. Our existing security page describes the simulation platform’s practices and provides a vulnerability reporting contact.
Where this stands today
This page states our approach and intended standards. It does not assert a completed frontier safety evaluation or certification. Future model-specific evaluations and deployment decisions will need their own evidence.