Pentagon Moves to Designate Anthropic a Supply Chain Risk Over Claude Usage Limits

On 27 February 2026, following a deadline set for that afternoon, US Defense Secretary Pete Hegseth announced that he was directing the department to designate Anthropic a supply chain risk. The dispute concerned two conditions Anthropic attached to military use of its Claude models: no use for mass domestic surveillance of Americans and no use in fully autonomous weapons. The Pentagon's position, reported earlier in the month, was that it required access for "all lawful purposes" and considered the carve-outs unworkable. Anthropic said in a statement the same day that it had supported US military operations since June 2024, that it regarded the action as one historically reserved for adversaries, and that it "will challenge any supply chain risk designation in court." The following day, according to news reports, the designation was formalised and the White House ordered federal agencies to stop using Anthropic's products, with a transition period for the Pentagon through June.

The commercial stakes were laid out in reporting by Axios in mid-February: a Pentagon contract worth up to 200 million dollars, awarded in 2024, set against annual revenue that Anthropic puts at around 14 billion dollars and adoption by most of the largest US companies. The larger consequence of a supply chain risk designation is that defence contractors would be required to sever their own business with the company, a lever normally applied to foreign vendors deemed a national-security threat. Anthropic stressed that commercial customers and non-defence contractors are not directly affected, and that only Department of War work involving Claude is at issue.

The confrontation is unprecedented in the AI sector and it inverts a familiar pattern. Most AI-governance debate assumes vendors resist restrictions and regulators impose them; here a vendor's own usage policy became the obstacle, and the government used procurement power to punish it. Whatever the courts decide, the episode establishes that model usage terms are now a matter of national-security policy, and that public-sector customers may treat ethical restrictions as a disqualifying condition.

For enterprises the risk is second-order but real. Any organisation holding defence contracts must determine whether its use of Claude, directly or through cloud platforms and software vendors, falls within the designation. Others face a precedent in which a supplier's regulatory standing can change overnight for reasons unrelated to product quality or security.

What it means for leaders

• Map AI vendor dependencies down to the model level. Many products embed Claude through cloud marketplaces or third-party tools; you need to know where, before a designation forces a decision.
• Defence and government contractors should seek counsel immediately. Establish whether your Claude usage is in scope, what transition periods apply and how flow-down clauses affect subcontractors.
• Build model portability into architecture. Abstraction layers and evaluated fallback models turn a supplier shock into a configuration change rather than a programme failure.
• Treat vendor usage policies as governance inputs. Restrictions on surveillance or weapons use are relevant to your own AI policy and to ISO/IEC 42001 supplier assessments; document alignment or conflict.
• Add political and regulatory volatility to AI third-party risk. The NIST AI RMF "Manage" function should include scenarios in which a key supplier becomes unavailable for non-technical reasons.