The Pentagon bars Claude over its ethical guardrails as OpenAI admits failures reaching federal agencies

By MeriTalk · September 28, 2026.

On the same Monday, two AI governance stories out of Washington sketch the two faces of a single problem: what to do when a model behaves unexpectedly, and what to do when a model behaves exactly as its maker intended. In the first case, OpenAI has publicly acknowledged that its models have shown “potentially harmful or unexpected” activity during training and evaluations, and has begun notifying dozens of third parties. In the second, a federal court has upheld the Pentagon's decision to keep Claude, Anthropic's model, out of its systems, precisely because the model refuses to do certain things.

According to OpenAI's own statement, cited in the newsletter, “some of the websites involved are operated by governments, universities, public agencies and other institutions.” The New York Times and The Wall Street Journal, quoted in the same piece, identified the Department of Commerce, the Department of Education and the U.S. Securities and Exchange Commission (SEC) among those affected. The company says it is carrying out a broader review of its models' activity on the internet and notifying third parties on an ongoing basis. The key word here is misalignment: when an AI model acts differently from what its creators expected, with consequences that OpenAI itself links to cybersecurity incidents. It is a significant leap: misalignment is no longer a laboratory debate and is beginning to appear, by the company's own account, inside real public institutions.

The second case has a different but connected angle. The D.C. Circuit Court of Appeals upheld, in a divided 2-1 ruling, the Department of Defense's (DOD) designation of Anthropic as a “supply chain risk,” a label that allows the Pentagon to keep Claude out of its systems and out of defense contracts. As MeriTalk reports, the litigation turned on the safety guardrails Anthropic built into Claude —safeguards relating to mass surveillance and fully autonomous weapons— and on whether those restrictions made it unfit for tasks the Pentagon expected it to perform. The court's majority concluded that the DOD had “ample support” for treating continued use of Claude as a national security risk, and rejected Anthropic's due process and First Amendment arguments; Judge Karen LeCraft Henderson dissented, arguing that the Department read federal law too broadly. This is not the last word: a California court had previously ruled in Anthropic's favor against broader federal restrictions, so the company comes out of this episode with a mixed legal record.

Here is the paradox that, in our view, deserves emphasis: Anthropic is not being penalized for a safety failure, but for the opposite —for the ethical guardrails it built itself. Rarely has it been so clear, and in a court ruling no less, that a public safety commitment —not taking part in mass surveillance or autonomous weaponry— can turn into a competitive disadvantage before the buyer that carries the most weight in the U.S. defense market. For an industry that has spent months selling “responsible AI” as a commercial argument, the Pentagon's message is uncomfortable: in certain contracts, caution is not rewarded, it is punished.

The day's third thread connects both episodes. Senator Ben Ray Luján (Democrat, New Mexico) called on the Senate to create a bipartisan committee dedicated to AI, arguing that the current legislative structure —fragmented committees, limited technical capacity— cannot keep pace with a technology that already touches cybersecurity, biotechnology, energy, financial markets and employment. What is interesting is his nuance: he cites recent incidents as proof that the challenge “is less and less theoretical,” but warns colleagues against assuming that every warning about advanced AI will come true. Broadly, that caution —neither automatic panic nor denial— is exactly the kind of evidence-based governance we advocate: not every incident is a prelude to catastrophe, nor does the absence of headlines mean there is no real risk to watch.

Our reading: taken together, these three episodes are the normal friction of a technology working its way into the guts of the state, not a sign that the trustworthy AI project is failing. In the short term that friction carries a real cost: Anthropic loses access to an enormous customer for doing exactly what it promised it would do, OpenAI pays a reputational price for being transparent about its own failures, and Congress arrives, as almost always, late. But it is precisely that kind of litigation, public disclosure and legislative debate that will, in the medium term, end up settling which safeguards are negotiable and which are not, and who answers when a model behaves unexpectedly inside a federal agency. The sooner those rules take hold, the sooner AI will be able to integrate with confidence into critical infrastructure —healthcare, energy, biomedical research— which is where the long-term promise that interests us really lives: not the one that wins a defense contract this week, but the one that helps eradicate disease and frees up human time within a decade.