Details
- Anthropic says it has been working with the US government since June 12 to restore access to its Claude Mythos 5 and Fable 5 models after an export-control suspension affecting all users, including foreign nationals inside the US.
- The company describes Mythos 5 as its strongest cybersecurity model, previously restricted over national security concerns tied to jailbreak techniques that could expose powerful vulnerability-finding capabilities.
- According to the latest update, US officials have now notified Anthropic that Mythos 5 can be redeployed in a controlled way to a defined set of US organizations that operate and defend critical infrastructure, such as utilities, telecoms, and core internet services.
- Access to Fable 5 remains constrained pending further government review and additional safeguards, as regulators weigh the risks of broader availability of offensive and defensive cyber capabilities embedded in these models.
- The redeployment appears to build on pre-release programs where a small number of security teams and agencies used Mythos 5 to identify vulnerabilities in long-standing software stacks, demonstrating both its utility for defense and the reasons for heightened oversight.
- Anthropic is positioning this partial restoration as a step toward more stable, rules-based access to high-capability AI for defenders while continuing to negotiate with US authorities over export controls, usage constraints, and monitoring obligations.
- The move suggests Mythos 5 will effectively function as a specialized, high-assurance tool for vetted US critical infrastructure customers, separate from Anthropic’s general-purpose Claude models that were never affected by the suspension.
Impact
Allowing Mythos 5 back into production for vetted US critical infrastructure operators signals that Washington is willing to tolerate tightly controlled deployment of frontier cyber-capable models while broader export and access rules are still being written. It narrows, but does not close, the gap with rivals whose advanced models remain more widely available, and underscores that leading AI vendors may increasingly run dual-track product strategies: heavily regulated, high-assurance models for sensitive sectors alongside more broadly accessible general-purpose systems.
