Trump Administration Blacklists AI Firm Anthropic

A striking contradiction now defines the U.S. government’s stance toward a leading artificial intelligence firm. While the Pentagon wages a legal battle to blacklist Anthropic as a national security risk, top economic officials are actively urging the nation’s largest financial institutions to adopt the company’s newest AI model for their own cyber defenses. This split highlights a deeply fragmented policy approach to a technology that is rapidly reshaping both security and commerce.

Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell recently convened executives from JPMorgan Chase, Goldman Sachs, Citigroup, Bank of America, and Morgan Stanley. Their directive was clear: test Anthropic’s Mythos AI model to uncover cybersecurity weaknesses within their vast financial networks. This recommendation stands in direct opposition to actions taken by the Defense Department just months earlier. Defense Secretary Pete Hegseth had formally designated Anthropic a supply chain risk, a move that bars military contracts and instructs defense contractors to cease using its technology. That designation followed Anthropic’s refusal to remove two core safety restrictions from its AI systems: a ban on use in fully autonomous weapons and a prohibition against mass surveillance of American citizens.

The capabilities of the Claude Mythos Preview model are what precipitated this bureaucratic clash. Anthropic did not specifically train this frontier model for cybersecurity work. Instead, its ability to find software vulnerabilities emerged as a byproduct of general advances in code reasoning and autonomous operation. During internal evaluations, Mythos reportedly identified thousands of zero-day vulnerabilities, previously unknown flaws, across every major operating system and web browser. Deeming the model too powerful for a general release, Anthropic created Project Glasswing, a restricted access program. Roughly 50 organizations, including several major banks and tech giants like Amazon, Apple, and Google, can now test the tool. The company has committed substantial funding, including usage credits and direct donations to open-source security groups, as part of the initiative.

Some security experts have questioned the narrative of a model too dangerous to release, suggesting the controlled rollout may be as much a savvy enterprise marketing strategy as a safety measure. They note that claims of discovering thousands of critical flaws were based on a limited number of manual reviews and that many identified vulnerabilities were in older software or difficult to exploit. Regardless of the marketing, the model’s perceived potency has driven swift interest from the financial sector.

The Pentagon’s dispute with Anthropic escalated in February when Hegseth presented CEO Dario Amodei with an ultimatum: remove the safety guardrails or lose a $200 million defense contract. Amodei’s refusal triggered the supply chain risk designation and a presidential order for federal agencies to stop using Anthropic’s technology. The subsequent legal proceedings have yielded mixed results. One federal judge issued a preliminary injunction against the designation, criticizing its overreach, while an appeals court allowed the blacklisting to stand during litigation. The current result is that Anthropic is locked out of Defense Department work but remains available to other government branches and the private sector.

It is precisely into this gap that Bessent and Powell have stepped. Financial institutions are now exploring Mythos for vulnerability detection, fraud-risk flagging, and compliance workflow automation. Their rapid engagement is fueled by a defensive imperative: if an AI can find hidden flaws in common software, it can likely find them in proprietary banking infrastructure as well. The logic is to discover and patch these weaknesses before a hostile actor using similar technology can exploit them. This concern has spread internationally, with UK regulators at the Bank of England and the National Cyber Security Centre also scrambling to assess the risks Mythos has uncovered.

The situation reveals an uncomfortable structural problem in the administration’s AI governance. One arm of the government condemns a company for adhering to ethical guardrails, labeling it a threat, while another promotes that same company’s technology as vital for protecting the core of the financial system. For Anthropic, the contradiction is strategically beneficial. Each bank that integrates Mythos embeds the company deeper into the nation’s critical infrastructure, making the Pentagon’s supply chain risk argument appear increasingly disconnected from reality. For the administration, the episode demonstrates how policy driven by personal grievance rather than coherent strategy can lead to one hand blacklisting what the other hand is urgently deploying. The banks, following the guidance of their primary regulators, seem content to proceed regardless of the Pentagon’s objections.