U.S. Agencies Test Anthropic Models Despite Pentagon Blacklist

The Department of War (the Pentagon) designated Anthropic as a supply chain risk, effectively blacklisting the company from defense work, while federal civilian agencies have begun testing Anthropic's newest model, Mythos, and the White House is negotiating broader civilian access. The D.C. Circuit denied Anthropic's request for a stay, with the court stating "In our view, the equitable balance here cuts in favor of the government," even as a separate federal judge granted a preliminary injunction in a different case. Anthropic has sued the government and continues to press its legal challenge.

Anthropic markets Mythos as a frontier-capability model with strengths in multi-step reasoning, code analysis, and cyber threat detection. Practitioners should note these operational characteristics:

• •Mythos is being evaluated for automated vulnerability discovery, adversary-lifecycle analysis, and rapid triage of incident data.
• •Anthropic previously supplied Claude variants and Claude Opus 4.7 for multi-modal tasks; Claude was formerly available in the Pentagon's classified network before the designation.
• •Anthropic sought contractual usage constraints that would ban mass surveillance and autonomous-weapons use; defense officials rejected those contractual limits, triggering the breakdown in talks.

This is a junction of technology, procurement law, and national-security tradeoffs. Frontier models like Mythos deliver capabilities that civilian agencies judge essential for cyber defense and infrastructure protection, yet those same capabilities create dual-use concerns when procurement terms restrict military uses. The current split between the Pentagon and the White House underscores competing priorities: operational control and battlefield readiness on one side, and broader civilian agency demand and industrial policy on the other. The legal record is fragmented. The appeals court prioritized government discretion during active conflict, while other federal courts have flagged possible overreach, producing inconsistent injunction posture across jurisdictions.

Agencies already testing Mythos will produce early operational data on real-world strengths and failure modes relevant to model safety, red teaming, and prompt engineering for security tasks. Expect experiments focused on false positive rates in vulnerability detection, model hallucination patterns when asked for exploit construction, and safe-query gating for sensitive outputs. The White House meetings led by CEO Dario Amodei and Chief of Staff Susie Wiles, described as "productive and constructive," signal potential policy carve-outs or standardized deployment guardrails that could set de facto federal practice for handling high-risk models.

Legal appeals could resolve the scope of the Department of War designation and determine whether usage-based contractual limits are enforceable. Operationally, look for published federal test results or briefings from the Commerce Department's Center for AI Standards and Innovation and for standardized guidance on safe cyber use cases. Also watch whether the administration pursues a centralized clearance or differentiated access model that separates civilian and defense deployments.

Frontier AI capability is colliding with procurement and safety policy. Practitioners should treat ongoing federal tests as a source of operational telemetry and an early signal for how the U.S. government will operationalize governance, access controls, and risk mitigation for high-impact models.