US Broadens Anthropic AI Ban, Extending Restrictions to Major Federal Agencies

The expansion, announced in a PYMNTS.com report, adds seven more departments—including the Department of Defense, the Department of Energy, and the National Security Agency—to the original list of agencies barred from deploying Anthropic’s Claude models. The move follows a 2023 directive that prohibited the use of “high‑risk” generative‑AI systems in agencies handling classified or mission‑critical data, a policy the administration has now tightened amid growing concerns over model transparency and data leakage (PYMNTS). Federal officials say the broadened ban is intended to give agencies more time to evaluate the security posture of Anthropic’s offerings before any production rollout.

The timing coincides with Anthropic’s aggressive product push, highlighted in a Tom’s Hardware story that detailed the launch of a new Claude‑based tool capable of generating legacy COBOL code. The demonstration, which produced a functional 67‑year‑old banking routine, sparked a sharp reaction in the financial‑services market, sending IBM’s stock down 13% after analysts warned that the capability could undercut demand for legacy‑system maintenance contracts (Tom’s Hardware). Anthropic’s marketing material frames the feature as a “code‑generation assistant for mainframe modernization,” but the rapid market impact underscores why regulators are wary of unvetted AI in critical infrastructure.

Further complicating the picture, The Register reported that Anthropic recently authored a 23,000‑word “constitution” for Claude, outlining internal guardrails and usage policies. The document, released publicly as part of the company’s transparency initiative, attempts to codify how the model should respond to disallowed content and mitigate bias (The Register). Critics argue that such self‑imposed rules may not satisfy federal security standards, especially when the models are integrated into environments handling classified data. The government’s ban therefore reflects a broader skepticism that private‑sector governance can replace formal oversight in high‑stakes settings.

Ars Technica added another layer by describing Anthropic’s new “Cowork” platform, a Claude‑Code‑like environment designed for general‑purpose computing tasks. Cowork promises to let developers write, test, and debug code across multiple languages within a single AI‑driven workspace (Ars Technica). While the feature showcases Anthropic’s ambition to compete with Microsoft’s Copilot and Google’s Gemini, it also raises the specter of “model‑in‑the‑loop” security vulnerabilities, where code suggestions could inadvertently embed malicious patterns or expose sensitive data. Federal IT leaders have cited these risks as part of the rationale for extending the prohibition to agencies that manage national security and critical energy infrastructure.

In sum, the broadened ban reflects a convergence of regulatory caution and market turbulence. As Anthropic rolls out increasingly sophisticated code‑generation tools, the federal government is signaling that adoption will only proceed once the company can demonstrate robust, auditable safeguards that meet the stringent requirements of the nation’s most sensitive agencies. The next few months will likely see Anthropic intensifying its compliance efforts while competitors watch to see whether a clearer regulatory pathway emerges for generative‑AI in government‑grade applications.