Punishing AI safeguards sets a costly precedent: 17 former senior officials back Anthropic against the Pentagon

By MeriTalk (via LinkedIn) · October 23, 2026.

Seventeen former US national security officials are asking a federal appeals court to reconsider its decision upholding the designation of Anthropic as a “supply chain risk,” according to MeriTalk's The Wrap newsletter. The signatories include former Director of National Intelligence Avril Haines and former national security advisers Jake Sullivan and Susan Rice.

According to MeriTalk's account, the dispute began when Anthropic objected to the Pentagon using its Claude model for mass surveillance and fully autonomous weapons. After negotiations failed, Defense Secretary Pete Hegseth chose to designate the company a supply chain risk. The former officials argue, again according to that source, that the Pentagon made the decision before it had a justification, and warn that the precedent could push advanced AI capabilities away just as the US competes with China. These are their arguments; the material does not include the Department of Defense's position or the court's reasoning.

It is worth gauging what this is: a politically weighty supporting brief, not a change in outcome. The court has already ruled one way, and the text does not say whether it will agree to reconsider.

Our take: what matters here is not Anthropic but the rule being set. If a company that places limits on specific military uses receives a label designed for suspect suppliers, the incentive for the rest of the industry is clear: don't ask uncomfortable questions. That would be bad in the short term, because the highest-risk uses (mass surveillance, lethal autonomy without a human) would be left without any counterweight from inside the labs themselves.

There is also a nuance that Anthropic's defense should not obscure: a state has the right to decide whom it buys from and on what terms, and it can argue that a supplier should not veto lawful uses. The underlying debate is who sets those limits. Our position is that they should be set by public, auditable rules, not by the whim of a supplier or a minister; until such rules exist, contractual safeguards are the only brake available, and penalizing them is counterproductive.

In parallel, the same newsletter covers another item consistent with this debate. Representatives Pat Harrigan (Republican) and George Whitesides (Democrat) introduced the AI Cyber Defense Act on October 22. It would task NIST (the National Institute of Standards and Technology) with developing voluntary standards against “distillation attacks,” a technique by which a third party extracts capabilities from an advanced model by querying it. It would also create a cybersecurity roundtable with tech companies and operators of power grids, hospitals and financial systems, to be set up within 180 days of passage. It is a bipartisan bill built on voluntary standards: useful, but, as with other voluntary initiatives, the decisive question will be what happens if someone doesn't comply.

Finally, the newsletter quotes officials from the NIH, the State Department and the Library of Congress who are looking to move AI beyond the pilot phase. One official says the State Chat tool saves about two hours per user per week; that figure comes from a single statement, not an independent measurement.

The thread tying it all together: AI is already at the heart of public administration, and governance (who sets limits, who enforces them, who pays for setting them) is lagging behind deployment. In the long run, the value of this technology will only materialize if that governance is built on clear rules rather than through ad hoc designations.