Pentagon–Anthropic dispute splits courts over AI safeguards | Magica
Pentagon’s Anthropic fight splits courts over military AI control
Editorial Team
••📖7 min read
The Pentagon treated Anthropic’s military-use safeguards as a supply chain threat. Later court rulings diverged under different statutes, while the dispute exposed how difficult it was to replace Claude in classified systems.
Courts reached different conclusions under separate statutes governing the Pentagon’s exclusion of Anthropic.
Anthropic disputes both the reliability of fully autonomous weapons and the Pentagon’s account of its ability to interfere with deployed systems.
The Pentagon’s March filing said replacing Claude required a 180-day transition, despite other suppliers being under contract.
The Pentagon’s effort to remove Anthropic, developer of the Claude family of AI models, produced conflicting rulings: a California district court rejected its supply chain designation in August, while a D.C. Circuit panel upheld a separate exclusion authority in September. The subsequent court decisions changed the legal position without settling the underlying fight over who controls military AI safeguards.
The dispute goes beyond the government’s choice of supplier. It tests when a vendor’s disclosed limits on military uses can justify excluding its technology from defense supply chains, including work delivered through other contractors.
The D.C. Circuit majority distinguished the narrower Section 3252 definition from Section 4713. Official court opinion, September 25, 2026, page 30. Source: D.C. Circuit opinion, September 25, 2026.
How safeguards became a security allegation
The Pentagon’s March 19 opposition to Anthropic’s request for a preliminary injunction argued that the company’s continuing access to model weights and safeguards could let it change a deployed system’s behavior without military consent. Officials warned that vendor interference during an operation could endanger lives.
Emil Michael, the Pentagon’s under secretary for research and engineering and chief technology officer, helped develop that assessment. A former Uber executive, he oversees the organization under which the Chief Digital and Artificial Intelligence Office, or CDAO, was placed in 2025. The government’s filing says CDAO’s mission is to accelerate military adoption of AI and that it took responsibility for AI supply chain risk assessments.
OpenAI reports sharply lower reasoning-extraction rates in internal GPT-6.2 tests and backs sanctions against illicit foreign distillation. Its own research and rival Anthropic’s experience show why protecting hidden reasoning leaves capability transfer and customer costs unresolved.
Editorial Team
The filing alleged that Anthropic had demanded an approval role in operational decisions and questioned a military use already permitted under its contract. It also cited the company’s public conduct during negotiations as grounds for distrust. These were the government’s allegations supporting exclusion, rather than findings by the court hearing that motion.
Anthropic contested the premise. In a March 20 declaration described in a subsequent report, its head of public sector, Thiyagu Ramasamy, said disabling or altering the technology in the manner the government alleged was not technically possible. The company argued that the government’s case rested on technical misunderstandings and concerns never raised during months of negotiations.
That rebuttal matters: a contractual refusal, a model’s built-in limitations and a supplier’s ability to change a military deployment are different mechanisms. The government’s March brief forecast possible sabotage; it did not document an actual shutdown or act of sabotage.
Anthropic had accepted military work, with limits
The department had chosen Claude despite Anthropic’s longstanding public safety policies. The March filing, citing Anthropic’s declarations, says the government first accessed Claude in March 2025, contracted with Anthropic that July and granted the company a Top Secret security clearance during 2025. Claude was accessed through government contractor Palantir Technologies, which partnered with Anthropic, and through a separate Anthropic agreement. National-security users operated under a government-specific addendum to its usage policy.
In its February 26 statement, Anthropic said Claude supported intelligence analysis, modeling and simulation, operational planning and cyber operations. It accepted the department’s authority to make military decisions and denied objecting to particular operations or imposing ad hoc restrictions.
Its objections were more specific than a general ban on weapons work:
Mass domestic surveillance: Anthropic argued that AI could combine scattered information about Americans into comprehensive profiles at scale, and that existing law had not caught up with those capabilities.
Fully autonomous weapons: It said current frontier AI was insufficiently reliable to select and engage targets without humans involved. It supported partially autonomous weapons and acknowledged that fully autonomous weapons might eventually be important for national defense. It said it had offered reliability research with the department, which had not accepted.
Anthropic said those two uses had never been included in its contracts. It also offered to support an orderly transition if the Pentagon chose another provider, keeping its models available on the terms it had proposed for as long as required. Anthropic said the department had threatened to use the Defense Production Act to force the safeguards’ removal, alongside threatening exclusion.
The department’s January 6, 2026 AI strategy instead directed procurement officials to incorporate language permitting any lawful use into AI contracts within six months, according to the government’s filing. Defense Secretary Pete Hegseth personally delivered the request to Anthropic’s chief executive on February 24.
The Pentagon said it had no interest in mass surveillance of Americans or autonomous weapons operating without human involvement. Its objection was to a private company imposing limits beyond the law. Anthropic’s position was that legality alone did not establish adequate safeguards or technical readiness.
The terms also mattered commercially. The government’s filing quoted Anthropic’s argument that dropping its restrictions would compromise its competitive identity, erode trust and impair its ability to recruit and retain staff. Its safety commitments were part of the business the Pentagon had selected, as well as the source of the later disagreement.
Contracts for Anthropic, Google, OpenAI and xAI each had ceilings of $200 million. The Pentagon’s March filing described a 180-day transition period. Ceilings do not establish spending or replacement cost. Graphic: Magica. Source: Congressional Research Service.
Other contracts did not make deployments interchangeable
The department had contracted with competing suppliers. In July 2025, it awarded contracts to Anthropic, Google, OpenAI and xAI for up to $200 million each, according to the congressional research briefing.
Google Public Sector, Google’s government-facing cloud business, announced its $200 million-ceiling contract on July 14, 2025. It offered AI models, computing infrastructure and cloud services. Google also said its distributed cloud products had obtained authorizations for sensitive classified applications. Alternative infrastructure therefore existed; that did not establish that another model could immediately take over Claude’s deployed work.
These contract ceilings are maximum values, not evidence of spending, comparable model prices or the cost of replacing Claude.
In March, the government’s filing, citing Anthropic, described Claude as the department’s most widely deployed AI model and the only one embedded in its classified systems. Michael’s declaration said immediate removal was technically and operationally infeasible, particularly during active operations. The Pentagon allowed 180 days to migrate to alternatives without impairing readiness.
Anthropic argued that continued classified use undermined the security-risk designation. The government answered that the transition period reflected operational necessity. Neither position established that another vendor could deliver an equivalent deployment immediately, or quantified the migration’s cost.
The district court vacated the Section 3252 designation; the D.C. Circuit later upheld exclusion under Section 4713, a separate authority. Graphic: Magica; source: Congressional Research Service. Source: Congressional Research Service.
What further review must distinguish
President Donald Trump directed federal agencies to stop using Anthropic technology on February 27, with a six-month phaseout, according to the March filing. Hegseth’s formal determination followed on March 3; Anthropic sued on March 9.
The original designation relied on 10 U.S.C. § 3252, a procurement authority originating in legislation addressing threats such as malicious code and supplier sabotage. The government argued that it covered domestic suppliers and excluded Anthropic from covered national-security procurements as a contractor or subcontractor. It said unrelated private business remained permissible, while Anthropic alleged damage to customer confidence and partnerships.
The government’s brief also acknowledged objections from Microsoft and employees of OpenAI and Google that the designation introduced uncertainty harmful to innovation and competition. The government argued that national security outweighed those concerns.
The later rulings must be distinguished. As summarized in the October 5 congressional briefing, the California district court on August 27 found the Section 3252 designation unlawful and the challenged actions retaliatory, vacating the designation and blocking its enforcement. The D.C. Circuit’s September 25 ruling concerned 41 U.S.C. § 4713, under the Federal Acquisition Supply Chain Security Act of 2018.
The appeals majority upheld exclusion under that separate statute. It cited encoded restrictions that had prevented Claude from completing government-requested tasks and uncertainty over an overseas operation. It also rejected Anthropic’s constitutional challenges. The dissent argued that the statute required intentionally subversive and deceptive conduct, rather than the majority’s broader reading of manipulation.
Further review would have to distinguish disclosed product limitations from covert interference, and ordinary vendor selection from exclusion under special national-security powers. The practical question is equally concrete: what technical control does the supplier retain over a deployed model, and can the government demonstrate that the resulting risk warrants exclusion?
Alberta proposes 600-metre setbacks for large data centres and expanded utility oversight, but exemptions and rules still to be written will determine how much protection communities receive.
President Donald Trump backed Commerce Secretary Howard Lutnick as lawmakers scrutinize alleged delays in restricting a suspected route for AI chips to China. A new congressional inquiry also questions Nvidia’s influence on an existing investigation.
Amazon’s month-old Echo Time is $50 off for Black Friday. Its touchscreen, home hub and optional sleep-aware alarm cost $90 more than the reported Echo Spot price, while Alexa+ access and evidence of better waking remain key buying questions.
Merck will pair its experimental data with Isomorphic Labs’ AI to seek small-molecule cancer drug candidates. The deal follows Isomorphic’s $5 billion financing, but its payments, target and development timetable remain undisclosed.
Thinking Machines Lab released a roughly 20GB open-weights model that reuses its middle layers for reasoning. It brings the Inkling family to local hardware, with competitive coding results but workload-dependent costs and preliminary evidence on oversight.
Apple’s first HomePod mini update since 2020 adds Mac audio and four-speaker surround, but Siri AI remains unavailable, core home-hub features carry over, and reviews disagree on playback reliability.
Treasury Secretary Scott Bessent says an unnamed Anthropic founder will announce a large gift to children's investment accounts within weeks, but the amount, recipients and connection to existing philanthropic pledges remain undisclosed.
Google says air cooling will limit water use at its Visakhapatnam AI hub, but environmental proceedings and residents’ accounts raise broader questions about land, electricity, public incentives and the evidence behind promised benefits.
Nvidia slipped behind Alphabet to fourth in U.S. market value during November 20 trading. The reversal follows a steep selloff despite strong reported revenue growth, with a permanent CEO still to be chosen.
Google will close the Nest app on January 25, 2027 and retire Assistant on home speakers and displays in August. Device transfers are automatic, but continued voice access requires accepting Gemini; basic assistance remains free.
Amazon says Cairn agents exploited a MONAI Label zero-day to take over a university server. Other attacks failed, and maintainers list no patched package release.
Federal Reserve Governor Lisa D. Cook sees early AI pressure on coding, translation and young workers, but says a solid overall labor market and persistent inflation still call for tighter monetary policy.
President Donald Trump threatened a tax on U.S. companies building data centers abroad without specifying how it would work. The proposal raises questions for an AI export program that has received 78 applications and for domestic construction constrained by power and community opposition.
Donald Trump said only older chips should be sold to China while discussing a broader deal with Beijing. His remarks leave unresolved which products would qualify as Washington tightens access to American AI hardware and overseas manufacturing.
Treasury Secretary Scott Bessent expects no change to the US policy permitting limited H200 sales to China despite sparse orders. Nvidia’s earlier filing shows Chinese restrictions and tariff costs also obstruct sales, while the company warns that exclusion helps competitors.
AI infrastructure stocks fell on November 20 amid cloud financing concerns and a reported NAND demand warning. Customer contracts support the buildout, but IREN’s filing shows the conditions and cash demands behind its expansion.
Jensen and Lori Huang’s foundation has committed $350 million to CoreWeave computing for research recipients. The purchase links charitable access to Nvidia’s commercial ecosystem, while pricing, allocation and supplier-selection terms remain unclear.
The mathematicians claim finite-time blow-up without external forcing, a significant advance if verified. Their repository omits the numerical certificate and its checking computation, and says the patched Lean package was not rebuilt.
A new investigation describes decades of alleged humiliation under Nvidia co-founder and former CEO Jensen Huang. He denies a former HR chief’s claims, while employee support and stock rewards complicate the account of his management.