A Federal Court Just Sided With the Pentagon Over Claude’s Guardrails
The D.C. Circuit ruled 2-1 that the Department of Defense lawfully branded Anthropic a national-security “supply chain risk,” a label that bars military agencies and defense contractors from using Claude. Judges Gregory Katsas and Neomi Rao formed the majority; Judge Karen LeCraft Henderson dissented, warning the decision lets the government weaponize contract demands into security designations.
The fight traces back to a 2025 contract where Anthropic refused to strip Claude’s restrictions on lethal autonomous weapons and mass domestic surveillance. When the Pentagon’s deadline for “all lawful purposes” access passed, Defense Secretary Pete Hegseth applied the risk label. The majority held that a supplier’s own “deliberate restrictions on functionality” can qualify as a supply-chain risk even when the motive is safety, not sabotage — a costly precedent for Anthropic just as it heads toward a expected $2 trillion IPO.
Anthropic Resumes Charging for Some Refused Requests
As of September 24, Anthropic bills standard API rates for requests refused by safety classifiers before any output is generated, but only in three of five refusal categories: bio (dangerous lab methods), frontier_llm (helping train a competing model), and reasoning_extraction (trying to expose internal reasoning). Refusals under cyber and general policy violations stay free.
Anthropic says the change targets attempts to “circumvent Anthropic’s safeguards at scale,” and picked these three categories because they show low false-positive rates as of this month. Translation for builders: hammering the API with prompts designed to probe these specific limits is no longer free reconnaissance.
Cache Diagnostics Graduate From Beta
Cache diagnostics on the Claude API no longer need a beta header. Developers can now opt in with a plain diagnostics object on Messages requests, and every response includes a diagnostics field back — useful for actually seeing why a prompt-cache hit did or didn’t happen instead of guessing.
Claude Quietly Breaks a Two-Year-Old Physics Record
Anthropic researchers report that Claude autonomously computed a nine-loop scattering amplitude in N=4 super Yang-Mills theory — a toy model physicists use to test calculation techniques — one loop past the record Lance Dixon’s team set in 2023. Claude used two independent methods, a bootstrap approach and an indirect form-factor check, working mostly unsupervised in Python and SymPy.
The bootstrap run alone cost about $100 in compute, equivalent to roughly a week on 96 CPUs; the full effort ran $1,000–$2,000. Dixon called the setup “very fragile” — one mistake and “it all crashes down like a failed soufflé” — which is what makes an AI holding it together for a week notable. Physicist Matt von Hippel put it simply: “there is more low-hanging fruit out there than you’d expect.”
The Same Guardrails That Cost Anthropic the Pentagon Are the Point
These three stories are really one story. A federal court just confirmed that Anthropic’s safety limits — no autonomous kill decisions, no mass surveillance — are expensive: they’re the reason a $2 trillion IPO candidate is locked out of defense contracts. The same week, Anthropic published proof that a tightly constrained Claude can still run circles around a genuinely hard problem with almost no supervision, and quietly tightened the screws on people trying to probe exactly where its safety classifiers bend.
None of that wins back the Pentagon. But it is Anthropic’s answer to the obvious question its critics keep asking: if you won’t sell an unconstrained model, what's the constrained one actually good for? A nine-loop amplitude a human team needed weeks to match is one answer. Whether investors pricing that IPO care more about the missing military revenue or the safety brand it protects is the tension Anthropic is now running straight into, in public, on purpose.