Anthropic CEO discusses Mythos access with the White House after Pentagon clash

The US government has opened direct talks with Anthropic about access to Mythos, a model that demonstrated in tests the ability to find and exploit thousands of previously unknown cybersecurity vulnerabilities. On April 17, CEO Dario Amodei met at the White House with chief of staff Susie Wiles and Treasury Secretary Scott Bessent, and the administration described the conversation as introductory, productive and constructive. For Washington, this is no longer an abstract dispute over AI regulation but an attempt to gain control over a tool that is too important for cybersecurity to remain outside the government perimeter.

The conflict between Anthropic and the Pentagon began in late February. Defense Secretary Pete Hegseth demanded that the military get virtually unrestricted access to the company's AI models for any lawful purpose, including autonomous weapons systems and domestic surveillance. Amodei refused, arguing that the models are not yet reliable enough for combat use and that US legal frameworks are not keeping pace with the risks of mass surveillance.

In response, the Pentagon designated Anthropic a supplier posing a national security risk, effectively shutting it out of government contracts. In early March the company filed two lawsuits against the administration, but on April 8 an appeals court lifted the temporary judicial protection. Anthropic has since remained outside Defense Department contracts, although it can formally work with other federal agencies.

The paradox is that just ten days after losing the appeal, the company unveiled Mythos. According to internal tests, the model was able to find and exploit thousands of zero-day vulnerabilities in major operating systems and browsers, including bugs that security specialists had missed for decades. When tasked with building a working exploit, the system succeeded on the first attempt in more than 83% of cases. Mythos also became the first AI model to complete a full 32-step simulation of an attack on a corporate network without human involvement. The UK's AI Security Institute assessed it as a markedly more powerful tool for offensive cyber operations than all previously tested models.

That is why Anthropic did not open Mythos to the general public. Instead of a public release, the company launched Glasswing, a controlled-access program, and admitted about 40 vetted organizations to the model, among them Amazon Web Services, Apple, Google, Microsoft, Nvidia and JPMorgan Chase. The program's goal is to find vulnerabilities before attackers do and patch them before real attacks occur. Anthropic also pledged up to $100 million in credits for using Mythos and another $4 million for open-source security projects. For the company, this continues the same logic of restraint that led to the falling-out with the Pentagon: release an extremely powerful tool only under strictly controlled conditions.

Now the parties' interests have begun to overlap. The Treasury wants to use Mythos to find weaknesses in its own systems. Certain intelligence community bodies and CISA are already testing the model, and the White House Office of Management and Budget is preparing a regime under which civilian agencies could work with a restricted version of the system.

Resolution matters to Anthropic as well, but not because of defense revenue. Its annualized revenue is estimated to have already reached $30 billion, investors have offered a valuation of $800 billion, and the company is exploring an IPO. The problem lies elsewhere: the supply-chain risk label hurts its reputation in the enterprise segment and creates uncertainty for customers that work close to the government.

There is added pressure from outside the US. In the coming days Anthropic plans to give several British banks access to Mythos and, at the same time, expand its London office to 800 employees. On April 15, Bank of England Governor Andrew Bailey called Mythos a cyber-risk factor, after which British regulators and the largest banks began emergency discussions. Canadian Finance Minister François-Philippe Champagne, speaking at IMF meetings, described the model as a source of risks whose scale cannot yet be precisely calculated. For Washington this creates an uncomfortable scenario: its closest allies could get a strategically important tool before the US federal government does.

Friday's meeting did not resolve the dispute and did not end the court fight, but it signaled a shift. Where the story once looked like a hard confrontation between the company and the Pentagon, the White House is now seeking a compromise: access to Mythos for defensive cybersecurity through civilian agencies, without Anthropic giving up its own safeguards. If such an arrangement works, it could become the first practical template for how states negotiate access to the most dangerous AI tools without dismantling the restrictions that motivated them in the first place.