Trump administration pushes Anthropic to banks while the Pentagon fights the company in court

Anthropic is at the center of a conflict that is rare in American technology policy: Donald Trump's administration has effectively pushed the company out of defense contracts, yet almost at the same time has begun promoting its most powerful AI tool to the country's largest banks. On April 13 it emerged that the US Treasury and Federal Reserve leadership had recommended that JPMorgan Chase, Goldman Sachs, Citigroup, Bank of America and Morgan Stanley test the Mythos model for finding cyber vulnerabilities. For Washington this is not just another pilot: it concerns a system that some agencies consider too risky and others already see as a useful layer of protection for critical financial infrastructure.

The paradox is that the Pentagon earlier designated Anthropic a supply-chain risk. The cause was the company's refusal to drop two core restrictions on its models: a ban on use in fully autonomous weapons and a ban on mass surveillance of American citizens. Pentagon chief Pete Hegseth then demanded revised terms of cooperation, and when Anthropic refused, the company was cut off from Defense Department contracts.

The conflict quickly reached the courts. One federal judge temporarily blocked the measures against Anthropic, but on April 8 an appeals court declined to halt the restrictions while the case proceeds. As a result, the company remains excluded from defense projects but can work with other government bodies, including the administration's economic bloc. Mythos stepped into exactly this gray zone.

According to Anthropic, the model was not specifically trained for cybersecurity; the abilities emerged as a byproduct of general improvements in code reasoning and autonomous operation. During testing, Mythos found thousands of zero-day vulnerabilities in major operating systems and browsers. Because of how sensitive these capabilities are, Anthropic did not release the model publicly and launched a closed program, Project Glasswing, for about 50 organizations. Participants already include Amazon Web Services, Apple, Google, Microsoft, Nvidia, Cisco, CrowdStrike, Palo Alto Networks and JPMorgan Chase. The company also pledged up to $100 million in usage credits and another $4 million in direct support for security-related open-source organizations.

The logic for banks is clear: if the model really can find vulnerabilities faster than in-house teams or ordinary scanners, it is better to use it for defense before similar tools reach attackers. Banks are reportedly testing Mythos not only to find holes in their infrastructure but also for fraud-risk detection and compliance automation.

Interest in the model has spread beyond the US. British regulators, the Bank of England and bodies responsible for financial system stability are already discussing what new threats and benefits this level of automated vulnerability discovery brings.

Skepticism exists too. Some security specialists doubt that the loud claims of thousands of critical findings have been sufficiently confirmed by independent verification. Others in the industry argue that the "too dangerous for public release" framing works both as responsible oversight and as a strong commercial argument for large corporate clients. But even setting the marketing aside, the reaction of banks and regulators shows that major institutions no longer see such models as a lab experiment but as a tool of strategic cyber defense that may change the rules of infrastructure protection.

The main takeaway for the market is that the dispute over Anthropic long ago outgrew the story of a single Pentagon contractor. The Mythos episode shows how hard it is for a government to punish a developer for refusing to weaken its safeguards while depending on its technology in critical infrastructure. If the largest US banks do build Mythos into their defense processes, the administration's position will look even more contradictory: supposedly too risky for the military sector, yet almost necessary for the financial system. This is no longer a private contract dispute but a signal of how quickly frontier AI models are becoming part of national security and financial stability.