Pentagon Shifts Classified AI Workloads to Rival Vendors

The Pentagon has confirmed that approximately 90 percent of its classified artificial intelligence workloads have been migrated off Anthropic’s platforms. Under Secretary of Defense for Research and Engineering Emil Michael stated that the remaining tasks are on track to be fully transitioned by the end of the month. This shift marks the completion of a fast-tracked effort to remove the company’s models from the military’s most sensitive networks.

The decision follows a breakdown in negotiations where Anthropic insisted on contractual safeguards restricting the use of its AI for mass surveillance of U.S. citizens and fully autonomous lethal weapons. The Department of Defense rejected these terms, arguing that defense software must remain available for all lawful purposes without such constraints. Consequently, the department designated Anthropic as a national security supply chain risk, a status that is now being contested in federal courts.

Strategic shift toward multiple vendors

Beyond the immediate legal dispute, officials cited a broader strategic flaw in relying on a single model provider. Michael noted that depending on one vendor puts warfighters in a suboptimal position if that provider fails to keep pace with rapid industry changes. By adopting a multi-model ecosystem, the department aims to ensure that diverse generative AI capabilities are available to support real-world operations across intelligence and warfighting networks, regardless of which company leads the frontier at any given moment.

The military has begun deploying models from other frontier vendors, including OpenAI and xAI, into its systems. Michael indicated that field users have responded positively to these alternatives, noting that different models offer distinct attributes depending on the specific mission requirements. This diversification is intended to provide flexibility and resilience in a technology sector that changes rapidly, potentially on a weekly basis.

Legal battles and security concerns

The transition is occurring alongside ongoing litigation. Anthropic was designated as a supply chain risk under two different laws, with one case adjudicated in the Northern District of California. However, the more significant legal battle is currently before the D.C. Circuit Court, a higher forum considered more oriented toward national security matters. Officials expect a ruling in the coming months, and until a preliminary injunction is granted or denied, the supply chain risk designation remains in effect.

According to reporting by GN technics/ai (en-US), the department is now integrating these alternative models into both classified systems and unclassified operational tools like the GenAI.mil platform. While the specific deployment methods and security levels vary by workspace, the overarching goal is to maintain operational capability while navigating the complex legal and ethical boundaries of modern warfare technologies.

Operational feedback from field users

Feedback from military personnel using the new tools suggests a mixed but generally positive reception of the alternatives. Michael stated that users operating on Palantir’s Maven system have expressed preference for newer versions of OpenAI’s ChatGPT over the previous Anthropic models. Additionally, some operators have found xAI’s Grok to be significantly faster, highlighting that different models serve different tactical needs.

U.S. officials with access to the secure government-tuned AI models recently came online with GenAI.mil. While they requested anonymity to speak candidly about their experiences, the general consensus aligns with the department’s public stance: the move to multiple vendors has been successful. The trade-off for this flexibility is a more complex management structure, requiring the military to oversee and integrate disparate systems rather than relying on a single, unified platform.