Anthropic Chief Discusses Mythos Access With White House Amid Pentagon Dispute

The meeting between Anthropic's CEO Dario Amodei and the White House apparatus shows that the dispute around AI can no longer be reduced to laboratory competition or another round of regulation. The issue is now about state access to a model that can find and exploit vulnerabilities at a level comparable to offensive cyber operations. On Friday, April 17, 2026, Amodei was scheduled to meet with White House Chief of Staff Susie Wiles to discuss access to Mythos.

This model was presented by Anthropic on April 7 as a universal system, but in tests it demonstrated far more alarming capabilities. The company claims that Mythos discovered thousands of previously unknown zero-day vulnerabilities in all major operating systems and browsers, including bugs that went unnoticed for decades. When asked to convert the discovered issues into working exploits, the model, according to Anthropic, successfully managed this in more than 83% of cases.

The company separately emphasizes that Mythos became the first model to complete a 32-step simulation of an attack on a corporate network from start to finish. This is why Anthropic refused a public release. Instead, the company launched Project Glasswing — a controlled access program for a limited circle of verified organizations.

Among the partners are named Amazon Web Services, Apple, Google, Microsoft, Nvidia, Cisco, CrowdStrike, JPMorgan Chase, and Palo Alto Networks, as well as several dozen other entities responsible for critical software and infrastructure. The idea is straightforward: if the model can find critical holes before malicious actors, it should first be given to those capable of quickly closing the discovered problems. To launch Glasswing, Anthropic allocated up to $100 million in credits for using Mythos and another $4 million in donations to open-source projects and security-related organizations.

This cautious approach became the cause of conflict with the Pentagon. The dispute has been ongoing since February: U.S.

Defense Secretary Pete Hegseth demanded that the military be given the broadest possible access to Anthropic's models for any lawful tasks, including scenarios involving autonomous weapons and internal surveillance. Amodei refused to weaken security restrictions. After this, the Pentagon assigned the company the status of a supply chain risk, which effectively closed its path to defense contracts.

Anthropic filed federal lawsuits attempting to challenge this decision, but on April 8, an appeals court revoked preliminary injunction protection. As a result, at the time of publication, the company remained outside Pentagon contracts, though it could still work with other federal agencies. The paradox is that the state, which punishes Anthropic for refusing to remove protective barriers, simultaneously seeks access to the company's most powerful model.

Interest in Mythos is shown by the U.S. Treasury Department, intelligence community structures, and other agencies that need a tool to find weaknesses in their own systems.

The White House is seeking a scheme by which federal agencies could use a controlled version of the model without completely removing restrictions. The story has already gone beyond the U.S.

: the British AI Security Institute conducted its own assessment of Mythos, and Bank of England Governor Andrew Bailey publicly called the model a new cybersecurity risk factor for the financial system. At the same time, Anthropic is preparing to expand Glasswing to some British banks, and this adds a geopolitical dimension to the situation: Washington's allies may gain access to such a tool before certain American agencies. The main conclusion is that advanced models are beginning to change the balance of power not in theory, but in real politics and cybersecurity.

Restrictions that were recently considered internal policy for AI companies are now becoming the subject of direct negotiations between corporations, militaries, and governments. The Mythos story shows: the next major conflict around AI will not only be about what the model can do, but also about who has the right to control its most dangerous capabilities.