On June 5, 2025, Anthropic quietly pulled back the curtain on Claude Gov, a bespoke version of its Claude family of large language models crafted exclusively for U.S. defense and intelligence agencies. Unlike its consumer-facing counterpart—which errs on the side of caution, flagging or flat-out refusing to process sensitive data—Claude Gov is tuned to operate in classified environments, “refusing less when engaging with classified information” and delivering richer context around defense- and intelligence-specific documents.

Anthropic’s announcement, published on its own newsroom on June 5th, states that these models are “already deployed by agencies at the highest level of U.S. national security.” Access, the company emphasizes, is strictly limited to government entities cleared to handle classified data. However, Anthropic stopped short of revealing when deployment first began—leaving industry observers to piece together the timeline from a handful of oblique references and insider chatter.

Behind the scenes, Anthropic says it leaned heavily on direct feedback from its government customers to shape Claude Gov’s capabilities. The result is a model pipeline that outperforms its civilian siblings in several key areas:

• Classified-material handling: Where consumer models balk, Claude Gov forges ahead, ingesting and reasoning over secret or top-secret documents without the usual refusal triggers.
• Domain-specific comprehension: From multi-page intelligence reports to cybersecurity logs, Claude Gov parses jargon-laden text with a fluency far beyond that of its public versions.
• Language and dialect proficiency: Recognizing that national security often hinges on understanding regional dialects and encrypted communications, Claude Gov brings enhanced support for languages critical to defense operations.
• Cyber-analysis acumen: Ingesting raw cybersecurity telemetry—malware signatures, intrusion alerts, network-flow data—Claude Gov can flag anomalies and suggest threat mitigations more effectively than a standard Claude model.

Anthropic insists that Claude Gov “underwent the same rigorous safety testing as all of our Claude models,” even as it loosens certain guardrails for classified use cases. That commitment to safety is baked into the company’s broader usage policy, which explicitly bars any user from employing Anthropic’s technology to produce, modify, design, market, or distribute weapons, explosives, dangerous materials or other systems designed to cause harm or loss of human life.

Yet, Anthropic quietly carved out contractual exceptions for carefully vetted government work at least eleven months ago. These carve-outs allow agencies to use Claude Gov for tasks ranging from strategic planning to threat assessment, provided that the work falls squarely within legal and mission parameters. Prohibited activities—such as disinformation campaigns, weapon design, censorship systems, and malicious cyber operations—remain off-limits, though Anthropic reserves the right to tailor these restrictions to each agency’s legal authorities.

The company’s stance reflects a balancing act familiar to any AI vendor courting government contracts: enabling powerful new applications while trying to head off worst-case scenarios. As Thiyagu Ramasamy, Anthropic’s head of Public Sector, put it in statements to Nextgov, “We’ve created a set of safe, reliable, and capable models that can excel within the unique constraints and requirements of classified environments.”

Anthropic isn’t alone in this race. In January 2025, OpenAI unveiled ChatGPT Gov, its own government-only service, reporting that more than 90,000 federal, state, and local employees had already used it to translate documents, draft policy memos, and spin up custom applications. Meanwhile, Scale AI struck a deal in March with the Department of Defense to develop an AI agent program for military planning and has since inked a five-year contract with Qatar to automate civil-service operations.

Beyond commercial players, long-standing defense-tech outfits are also doubling down on AI. Palantir’s FedStart program—designed to help software vendors navigate federal procurement—counted Anthropic as an early partner, facilitating the Claude 3 and 3.5 deployments on AWS for classified workloads. And in separate efforts, the Department of Energy’s National Nuclear Security Administration “red-teamed” Claude 3 Sonnet to ensure it couldn’t inadvertently divulge sensitive nuclear-weapons info, marking the first known test of a frontier AI model in a top-secret setting.

The use of AI by government agencies has a checkered history. Wrongful arrests tied to face-recognition errors, biased predictive-policing tools, and opaque welfare-eligibility algorithms have all drawn fire from civil-rights groups. Public protests—like those organized under the “No Tech for Apartheid” banner—have targeted Microsoft, Google, and Amazon over their military contracts in conflict zones.

Anthropic’s Claude Gov launch thus comes at a moment of heightened scrutiny. Critics argue that loosening safety filters—even in the name of national security—risks entrenching algorithmic biases or enabling misuses that could disproportionately harm vulnerable communities. Anthropic counters that its policy framework and in-house safety teams will help prevent such outcomes, though skeptics note that classified programs often lack the transparency needed for an external audit.