Anthropic applied AI Safety Level 3 protections to Claude Opus 4 in May 2025 because it could not confidently rule out the relevant risks. The decision was precautionary: the company had not established that the model crossed its capability threshold. Deployment protections targeted assistance with chemical, biological, radiological and nuclear weapons workflows, including classifiers that check inputs and outputs. Separate security controls sought to prevent theft of the model's weights, the numerical parameters needed to run it. ASL-3 therefore described the safeguards applied under uncertainty, rather than a conclusive finding about the model's dangerous capabilities. Activation rationale.
