CISA Deploys Anthropic AI to Hunt Government Software Vulnerabilities

CISA’s Attack Surface Evaluation team is utilizing the model to identify deep-seated vulnerabilities within government infrastructure. Sources familiar with the project confirmed the audits have already exposed a significant number of bugs, though specific details regarding the nature or severity of these findings remain classified. Neither CISA nor Anthropic have publicly commented on the scope of the software review.

This deployment highlights the shifting dynamics between Washington and the San Francisco-based developer. Despite receiving a formal supply-chain risk designation from the Pentagon in February—an action triggered by Anthropic’s refusal to remove safeguards against autonomous weapons and surveillance—the company’s technology remains highly coveted by intelligence agencies. The National Security Agency reportedly began testing Mythos in classified settings as early as April, valuing its precision in detecting cybersecurity exploits.

The relationship remains fraught with friction. Tensions peaked recently when the White House demanded strict usage restrictions on the public version of the model, Fable, leading to a temporary global shutdown. While a judge blocked the initial blacklisting of the firm in March, the government’s reliance on Anthropic’s tools continues to coexist with rigorous efforts to control how the company’s AI models are accessed by foreign actors.