Federal agency personnel have been conducting discreet evaluations of Anthropic’s new AI model, Claude Mythos, to judge its usefulness in identifying critical software vulnerabilities, according to people familiar with the matter. The assessments involve staff from several offices, including the Commerce Department’s Center for AI Standards and Innovation.
Those involved are examining the model for its reported ability to detect vulnerabilities that might escape human reviewers. The work is being carried out quietly, even though a presidential directive earlier this year ordered agencies to halt use of Anthropic’s technology following disputes over the firm’s position on military and surveillance applications.
Officials say the model’s technical strengths have prompted consideration for defensive cybersecurity purposes. The capability to find hard-to-detect flaws has been described by some involved as potentially important for strengthening cyber defenses, which has driven efforts to assess and, in some cases, deploy the technology despite the formal restriction.
Interest has extended to the legislative branch, where lawmakers and congressional staff have requested briefings on the system. Those inquiries reflect a shared concern that adversaries could develop comparable tools, increasing urgency among officials to understand the technology and its implications.
The White House has said it remains in dialogue with AI companies to address security risks, while attempting to balance those engagements with broader national security considerations. Beyond confirming ongoing interaction with industry, the White House statement did not provide additional operational detail on the agency-level evaluations.
This sequence of actions - quiet technical assessments, congressional briefings, and continued White House engagement - highlights a tension between a formal policy restriction directed at a particular vendor and the perceived operational need to understand and potentially use advanced AI capabilities for defensive purposes.
Contextual note: The information about agency testing comes from individuals familiar with the activity; agencies involved include the Commerce Department's Center for AI Standards and Innovation. The presidential directive to halt use of the technology was issued earlier this year and followed disagreements over Anthropic's stance on military and surveillance applications.