NSA Runs Anthropic’s Claude Mythos on Classified Networks

Two U.S. officials confirmed the National Security Agency is running Anthropic’s Claude Mythos Preview inside classified networks. The use comes while the Defense Department has designated Anthropic a supply‑chain risk and litigation between the Pentagon and the company continues.

Anthropic limited Mythos access to a small, vetted set of organizations because internal testing showed the model could identify critical vulnerabilities in widely used operating systems and web browsers. The company formed a controlled-access group called Project Glasswing that includes major technology firms and some financial institutions. Most member organizations are using the model to scan their own systems for weaknesses.

Details about how the NSA is applying Mythos were not disclosed. One source described broader use across parts of the intelligence community. Agency officials did not outline specific tasks or target systems tied to the deployment.

Negotiations between Anthropic and the Defense Department soured after the two sides reached an agreement in July 2025 to clear Claude for use on classified systems. The Pentagon later sought broader contractual rights to use the model “for all lawful purposes,” a demand Anthropic rejected. The company set two limits: the model would not be used to develop autonomous weapons or for domestic mass surveillance.

Defense Secretary Pete Hegseth designated Anthropic a supply‑chain risk in late February. A federal judge in California temporarily blocked the designation, and a D.C. appeals court denied Anthropic’s attempt to pause the blacklisting while the legal dispute proceeds. The matter remains before the courts.

Anthropic CEO Dario Amodei met at the White House on April 17 with Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent. Anthropic described the meeting as “productive.” The White House said participants discussed collaboration and shared approaches and protocols to address challenges from scaling powerful AI systems. An administration official noted that every federal agency except the Department of Defense is seeking access to Anthropic’s tools. Treasury and Federal Reserve officials have encouraged major banks to test Mythos to prepare for security threats.

Independent researchers reported reproducing several of Mythos’s cybersecurity findings using publicly available large language models, including recent versions from other providers and Anthropic’s less‑restricted models. Those researchers used public models to replicate exploit discovery examples that Anthropic had flagged in its internal tests.

The NSA and Anthropic did not provide additional comment on the agency’s reported use of Mythos. President Trump said he had “no idea” Amodei had been at the White House.