Anthropic's Claude 5.1 release continues the structure it introduced in June: a generally available model with classifier-based safeguards on cyber and other dual-use capabilities, and an approved-access variant for organisations such as critical infrastructure defenders, security vendors and government evaluators that need the full capability. The company describes the gating as the operational answer to the finding, since Project Glasswing, that frontier models can discover and exploit vulnerabilities at scale.
The release lands a month after Anthropic's post-mortem on the summer evaluation incidents and the containment measures it adopted, and it is the first major release under the voluntary pre-release framework that Executive Order 14409 asked the NSA and CAISI to build.
What to do: if your organisation needs cyber-capable models for defence, apply through the approved-access programmes rather than attempting to remove safeguards; if you deploy the general model, review the published safeguard documentation and test your own agents against it.