Anthropic's Mythos AI Found NSA Vulnerabilities, But Reports Overstated the Threat

When Anthropic disclosed Mythos in April, it sent a shockwave through the cybersecurity sector. The AI model was so effective at finding security vulnerabilities that the company limited its release to early testers, including the U.S. National Security Agency. This week, reports emerged that Mythos had uncovered multiple vulnerabilities within the NSA's own systems.

During a June 11 Senate hearing, Senator Mark Warner said Mythos had broken into "almost all of [the NSA's] classified systems, not in weeks, but in hours." He attributed the information to NSA Director General Joshua Rudd. The claim spread rapidly, fueling fears that even the world's most advanced cyberdefenses were vulnerable to AI-powered attacks.

Controlled Tests, Not a Breach

Subsequent reporting clarified the situation. According to federal officials cited by the New York Times, the tests were conducted in a tightly controlled digital environment that could not be replicated by external hackers. Mythos identified vulnerabilities but did not actually exploit them. The Economist's author later admitted on X that his portrayal was misleading, saying he quoted Warner to give a sense of Mythos' potency but should have added caveats.

• Mythos was first disclosed in April 2026 and limited to early testers including the NSA.
• Senator Warner's statement during a June 11 hearing sparked widespread concern.
• The tests were conducted in a controlled environment; no actual exploitation occurred.
• The Economist's author acknowledged the report was misleading in a June 24 X post.

Political Restrictions Complicate Access

The controversy comes amid a political battle over access to Mythos-class models. The Trump administration earlier this month invoked an obscure export control law to restrict access for foreign nationals to Fable 5, a publicly available Mythos-class model, and also targeted Mythos 5, which was limited to a small group. The move halted the NSA's internal tests with Mythos. Cybersecurity experts argued the ban would hamstring U.S. defenses and give adversaries an advantage.

The administration is now working with Anthropic to reinstate the NSA's access for limited national security purposes. The incident highlights the tension between harnessing AI for defense and controlling its proliferation. As Anthropic prepares for what is expected to be a historic IPO, the scrutiny over Mythos' capabilities and risks is unlikely to fade.