Activating AI Safety Level 3 protections

Anthropic announced that, as it launches Claude Opus 4, it has switched on the ASL-3 (AI Safety Level 3) deployment and security standards drawn from its Responsible Scaling Policy (RSP). The company describes the step as precautionary and provisional, because it has not yet established whether the model actually crosses the capability threshold that requires these protections. It explains that, unlike with earlier models, it could not clearly rule out ASL-3 risks as CBRN-related knowledge and abilities keep improving, though it did rule out needing the stricter ASL-4 standard and found Claude Sonnet 4 does not need ASL-3.

All prior Anthropic models had been released under the ASL-2 baseline, which includes training models to refuse dangerous CBRN requests and defenses against opportunistic theft of model weights. ASL-3 demands stronger defenses against both misuse and theft, suited to sophisticated non-state attackers. Anthropic says acting early lets it refine the protections before they are strictly needed, and that the RSP permits erring on the side of caution. If further evaluation shows the threshold was not passed, the protections may be loosened or removed.

The deployment measures aim narrowly at stopping help with extended, end-to-end chemical, biological, radiological and nuclear weapons workflows, especially through universal jailbreaks. They rest on Constitutional Classifiers that monitor inputs and outputs in real time, wider monitoring including a bug bounty program and threat intelligence partnerships, and quick remediation of discovered jailbreaks. On security, more than 100 controls protect model weights, including two-party authorization, change management, binary allowlisting and preliminary egress bandwidth controls that limit data leaving secure environments.

Anthropic says the question of which measures suit frontier models is unsettled, and it will keep improving them and cooperating with industry, users, government and civil society.