U.S. officials say the NSA has preview access to Anthropic's Mythos — a powerful AI model the company withheld from public release — and is using it to scan systems for exploitable flaws.

What Mythos is and why Anthropic kept it closed Anthropic introduced Mythos this month as a frontier AI model tailored for cybersecurity tasks. The company said the model's capabilities made unrestricted release risky because it could be used to design offensive cyber tools. Rather than a wide public launch, Anthropic opened a preview programme limited to roughly 40 organisations and publicly identified about a dozen of them. That limited rollout left most potential users and observers without access. Anthropic framed Mythos as powerful enough that open access could enable misuse, so the firm applied strict controls on who can run the model. Anthropic declined further comment on the model or on which other organisations have preview access. How the NSA reportedly fits in A report says the National Security Agency is among the undisclosed recipients of Mythos Preview and is using the model mainly to scan environments for vulnerabilities that attackers could exploit. That role aligns with the NSA's long-standing mission to identify weaknesses in networks and infrastructure. Key points: - officials describe such tools as useful for defensive work and for mapping threats before adversaries find them. - The NSA's reported use is described as scanning systems for exploitable flaws rather than offensive development. - Anthropic has publicly named only about a dozen preview organisations; most recipients remain private. Pentagon feud, court fights and a narrow thaw The dispute between Anthropic and parts of the U.S. military has both policy and legal dimensions. Some military branches have argued in court that AI models can pose national-security risks if used for mass surveillance or autonomous weapons development; Anthropic's refusal to make its Claude model broadly available to the Pentagon helped trigger that dispute. At the same time, some federal defence offices appear to be using Anthropic's tools for defensive work. That tension — where one agency warns of risk while another uses the same technology — has complicated official positions on access and oversight. Access controls, transparency and the list of users Anthropic publicly named about a dozen organisations with preview access but kept most recipients private. That selective disclosure aimed to balance the company's stated safety concerns against pressure for transparency from governments and watchdogs. The inclusion of national security agencies in a private access list raises questions about oversight and safeguards: agencies that get limited access can use models in ways the company intended — for example, scanning for flaws — but those same capabilities could be repurposed if controls are weak.

Related Articles

The White House called a recent meeting between Anthropic CEO Dario Amodei, White House chief of staff Susie Wiles and Treasury Secretary Scott Bessent "productive"; the company's limited preview programme and reported use by national security agencies, however, leave lingering questions about oversight and safeguards.

This article was created with AI assistance.