Detailed Analysis
The Trump administration has entered into a paradoxical policy posture toward Anthropic, simultaneously maintaining a national security blacklist against the AI company while actively encouraging major U.S. financial institutions to adopt its most powerful and controversial AI model. The Pentagon, under Defense Secretary Pete Hegseth, designated Anthropic a "supply chain risk" after the company refused to strip safety guardrails from its Claude-family models — restrictions that prevent the technology from being used to enable fully autonomous weapons systems or mass surveillance programs. Despite that designation, Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell have been urging executives at JPMorgan Chase, Goldman Sachs, Citigroup, Bank of America, and Morgan Stanley to evaluate Anthropic's **Mythos** model, which has demonstrated a capacity to autonomously identify thousands of previously unknown zero-day vulnerabilities in operating systems and browsers. Distribution of Mythos is being managed through a restricted channel called **Project Glasswing**, with access limited to approximately 50 organizations, and UK regulators are separately conducting their own assessments of the model's capabilities.
The legal landscape surrounding Anthropic's blacklisting remains unsettled and consequential. A Washington D.C. appeals court upheld the Pentagon's designation, affirming the government's authority to restrict contracting with the company on national security grounds. However, a federal judge in San Francisco granted Anthropic temporary relief, blocking portions of a presidential executive order and citing potential violations of First Amendment free speech protections — a notable legal theory suggesting that mandating the removal of AI safety restrictions could constitute compelled speech. Appeals remain pending in both proceedings, meaning the ultimate resolution of Anthropic's federal standing is far from determined. The company previously held a $200 million Department of Defense contract, a relationship that has since been severed by the blacklisting dispute.
Anthropic co-founder Jack Clark confirmed that discussions with the Trump administration regarding Mythos and future model deployments are ongoing, framing the blacklisting conflict as a "narrow contracting dispute" rather than a fundamental breakdown in the relationship. Clark emphasized that Anthropic remains committed to national security collaboration in principle, even as the company refuses to compromise on its core safety constraints. This framing reflects a deliberate strategic posture: Anthropic is seeking to preserve its position as a credible government partner without capitulating to demands that it views as incompatible with responsible AI development. The simultaneous rejection and pursuit of Anthropic's technology by different arms of the same administration suggests that Mythos's demonstrated capabilities — particularly in offensive cybersecurity — have proven too operationally valuable to forgo, regardless of the broader policy dispute.
The situation illuminates a deepening tension within U.S. AI governance between security utility and safety architecture. The administration's position implies that Anthropic's guardrails are acceptable, or at least tolerable, when the application is financial infrastructure defense, but unacceptable when they limit military and intelligence applications. This selective approach to AI safety standards risks establishing a fragmented regulatory environment in which the permissibility of safety restrictions is determined by the end-use sector rather than any coherent principled framework. For the broader AI industry, the episode signals that companies can expect government pressure to customize or remove safety measures for specific high-value use cases, creating competitive dynamics that may disadvantage safety-conscious developers relative to those more willing to strip constraints at the government's request.
The Mythos episode also represents a broader inflection point in how advanced AI capabilities intersect with critical infrastructure security. A model capable of autonomously identifying thousands of zero-day vulnerabilities at scale has transformative — and dual-use — implications: it can harden financial systems against cyberattacks, but the same capability deployed offensively could destabilize them. The fact that distribution is being tightly controlled through Project Glasswing suggests that even the administration acknowledges the danger of broad access to such a system, an implicit concession to the very logic of AI risk management that underpins Anthropic's guardrail philosophy. How the courts ultimately rule on the blacklisting dispute, and whether Anthropic ultimately reaches a formal agreement with the government over Mythos deployments, will likely shape the contours of AI governance debates well beyond 2026.
Read original article →