According to a Politico report based on leaked emails and anonymous sources, the Pentagon is exploring whether leading artificial intelligence models — including Anthropic’s unreleased Claude Mythos Preview — could be used for military purposes.
The report stresses that, in principle, the Pentagon is not allowed to use Claude Mythos Preview because Anthropic has not released the model. The tool has been judged potentially too dangerous, and tensions between Anthropic and the U.S. government have increased recently, accompanied by ongoing legal proceedings. As a result, the Pentagon officially does not use Anthropic products, and that restriction applies to its contractors as well, the Gizmodo piece notes.
New U.S. AI initiative and NSA involvement
Politico highlights that the U.S. AI initiative announced earlier this month includes not only the Pentagon but also the National Security Agency (NSA). The initiative’s aim mirrors part of Anthropic’s stated goal for the Claude Mythos Preview: to discover system vulnerabilities and fix them.
The task force is assessing how the Pentagon could safely deploy leading AI models across "all areas" of its missions and whether Silicon Valley-developed systems can be used within "high security classification" environments, including systems that protect the intelligence community’s most sensitive secrets.
Glasswing program and offensive uses
The report says that, under a program called Glasswing, the Pentagon would consider using AI models that possess hacking capabilities as weapons. Politico notes that Anthropic has previously indicated it is open to employing Mythos in offensive operations.
Overall, the outlet suggests the Pentagon is likely to "review" the possibility of employing Claude Mythos Preview, even though current legal and security constraints and the model’s unavailability remain limiting factors.
Why this matters
The issue extends beyond technical feasibility into legal, ethical, and security domains. Deploying high-capability models that could perform system intrusions or other offensive tasks would shift norms around military use of AI versus civilian use. Supplier relationships and ongoing legal disputes will also shape which products, and under what conditions, can be brought into operational military use.
