Safety Standards & Frontier Evaluations

Pre-deployment safety evaluations answer a regulatory question: what can this model do that would matter if it did? The evaluation categories that became policy-relevant:

• Dangerous capability evals — can it uplift bioweapon development, execute cyberattacks, or sustain autonomous operation? METR's time-horizon methodology measures the length of task a model can complete autonomously — a capability number regulators can track.
• Deception and scheming evals — does the model conceal capabilities or pursue hidden goals under test conditions?
• Refusal and robustness evals — does it decline what it should, under adversarial pressure?

These run before deployment (AISI pre-deployment agreements), in-house, and increasingly as conditions written into frontier labs' own policies — the eval result is becoming a deployment gate.