然而，政策回应已演变为争议。Anthropic 依据一项与国防部的合约供应专用版 Claude。公司据称要求限制大量监控与自主武器系统等用途，而五角大厦则主张对所有合法用途皆有权使用。经过公开对抗后，Pete Hegseth 将 Anthropic 指定为国家供应链风险，总统命令联邦机构在六个月内停止使用 Anthropic 的产品。现在已进入诉讼程序，且这一指定对于一家非外资控制、具有国安合作关系的美国 AI 公司而言十分罕见。文章将此举描述为对先前被视为安全相关基础设施伙伴的一种不常见的突兀转向。
这种冲突在战略与作业层面都矛盾。行政部门同时声称 Anthropic 对美国国家安全关键，而且报导指出 Anthropic 工具在美伊冲突中据称仍被使用。如果该指定持续，国防承包商可能会避免使用 Mythos，而本可参与 Glasswing 型漏洞猎捕的机构也可能因法律与政治风险而延后或拒绝合作。结果可能是协同揭露与回报参与率下降、修补周期变慢，尽管底层模型能力更强，整体防御仍会变弱。随著更多前沿模型获得类似的漏洞发现能力，该政策可能会适得其反，削弱企业与监管者之间的信任并提升长期安全暴露。
Anthropic PBC’s Claude Mythos Preview is described as a world-class cybersecurity model. Anthropic says it has already identified thousands of high-severity vulnerabilities, including in every major operating system and web browser, with some vulnerabilities remaining undetected for decades. Concerned by this impact, Anthropic withheld the model from public release, published a detailed model card, briefed U.S. authorities, and coordinated with major companies in Project Glasswing to patch flaws. The broader security-data context is that exposure counts change over time; the 2026 benchmark cited is as of April 17, 2026. In this framing, frontier-model scanning can shrink latent defect windows and accelerate remediation.
Yet the policy response has become a dispute. Anthropic supplies a specialized version of Claude under a Defense Department contract. The company reportedly sought restrictions on uses such as mass surveillance and autonomous weapon systems, while the Pentagon asserted authority for all lawful uses. After a public confrontation, Pete Hegseth labeled Anthropic a national supply-chain risk, and the president ordered federal agencies to stop using Anthropic products within six months. Litigation is now underway, and this designation is unusual for a U.S. AI company closely tied to national security. The article characterizes it as an abrupt reversal toward a partner previously treated as part of security infrastructure.
The contradiction is present in both strategy and operations. The administration simultaneously argues Anthropic is critical to U.S. national security, and reports indicate Anthropic tools are reportedly still used in the U.S.-Iran conflict. If the designation persists, defense contractors may avoid Mythos, while agencies that could otherwise engage in Glasswing-style vulnerability hunting may delay or refuse collaboration because of legal and political risk. The likely result is lower participation in coordinated disclosure and slower patching cycles, while overall defense worsens even as base-model capability grows. As more frontier models gain similar vulnerability-finding ability, the policy can become counterproductive, eroding trust between firms and regulators and increasing long-term security exposure.