Anthropic’s Mythos Found Flaws in Classified U.S. Systems Within Hours

When Anthropic’s Mythos model reportedly needed mere hours to identify vulnerabilities across highly sensitive U.S. government systems during a testing exercise with intelligence agencies, it added to the tension between the company and the government and complicated the governance debate.

A U.S. official cited by the Associated Press said Anthropic worked with those agencies to assess the model against classified systems. Mythos quickly identified certain vulnerabilities, although the official stressed that finding a flaw is different from exploiting it (AP News, 2026).

The testing formed part of Anthropic’s Project Glasswing, an initiative aimed at helping technology companies secure critical software against the risks posed by powerful AI systems. The project reflects a growing concern inside government: advanced models may soon become central to vulnerability research, exploit development and cyber operations.

Democratic Senator Mark Warner referenced the tests during a Senate Banking Committee hearing in June. Citing information attributed to General Joshua Rudd, head of the National Security Agency and U.S. Cyber Command, Warner said the tool had broken into “almost all” classified systems “not in weeks but in hours” (Yahoo News, 2026).

A model capable of quickly finding weaknesses in classified systems will inevitably be treated as a national security asset, not just a commercial AI product. The company has cooperated with U.S. agencies on security testing, yet tensions with the Trump administration have grown. Anthropic has raised concerns about how the U.S. military may use its AI systems, while the administration has moved to restrict access to some of its most advanced models.

The context is already tense as the disclosure lands amid growing friction between Anthropic and the U.S. government. Washington recently ordered the company to suspend access to its Fable 5 and Mythos 5 models for foreign nationals on national security grounds, only days after Anthropic launched them (CNBC, 2026). The move followed concerns from the Trump administration that the models’ cyber capabilities could create risks if made widely available.

The administration’s directive came shortly after President Donald Trump signed an executive order establishing a framework for the federal government to review national security risks posed by the most advanced AI systems before their public release. Participation by AI developers was described as voluntary, but the policy direction is clear: Washington wants more control over cyber-capable AI before it enters wider circulation. Anthropic disabled the models for all customers to comply with the directive, while saying it did not believe the government’s response was justified by the specific security concern it had raised.

That has triggered pushback from the cybersecurity community. More than 100 experts and industry leaders, including figures from companies such as Adobe and Nvidia, urged the government to lift the directive (Reuters, 2026). Many security professionals already use other foundation models and open-source systems for audits, training and defensive research. Removing access to one of the strongest tools could weaken U.S. defenders while doing little to stop adversaries from using alternatives.

If a model can quickly identify vulnerabilities, restricting access may reduce misuse. It may also deprive security teams of the same capability at a time when adversaries are improving fast.

Mythos makes the AI governance debate harder because its value cuts both ways. A model that can uncover vulnerabilities in sensitive systems can strengthen U.S. defense, but it can also help others find weaknesses faster.

That makes access control an incomplete answer. Restricting Mythos may reduce some misuse risk, but similar capabilities are already emerging beyond Washington’s reach. Since AI can find flaws in hours, defenders need to validate, prioritize and patch before the same weakness becomes leverage.