Anthropic's Restricted Mythos Model Exposes AI Cyberrisk Governance Gap
Anthropic restricted Mythos — not a dedicated cyber model but a general-purpose model restricted because it excels at everything, including cybersecurity — based entirely on its own internal risk self-assessment. Analysts note that OpenAI and Google are approaching the same capability threshold "soon," with no external regulation requiring them to apply equivalent restraint. Any competitor could release a Mythos-class model by simply assessing risk more permissively. The day Anthropic announced GPT-5.5-Cyber for defenders, Mythos illustrated the same problem from the supply side.
Why It Matters
The asymmetric governance risk is concrete: Anthropic's voluntary restriction may create competitive disadvantage with no corresponding safety benefit if peers release equivalent capability without restriction.