Dario Amodei's Pentagon emails show why Anthropic's military AI fight keeps escalating

Why it matters

Anthropic's Pentagon fight shows AI safety policy becoming a contract term, with founders negotiating directly against national-security buyers over model control.

Emails between Dario Amodei (@DarioAmodei) and Pentagon Undersecretary Emil Michael show how Anthropic's military AI negotiations hardened into a test of who gets to set the operating limits for frontier models inside national security systems, the Wall Street Journal reported Thursday.

Amodei, who co-founded Anthropic with Daniela Amodei in 2021 after serving as OpenAI's vice president of research, has built Anthropic around the claim that safety is a product requirement rather than a public-relations layer. The Pentagon fight forced that premise into a government contract. Anthropic wanted two exceptions to the Pentagon's demand for Claude access for all lawful uses: no mass domestic surveillance of Americans and no fully autonomous weapons.

The Journal's account lands one day after Anthropic restored access to Fable 5 and Mythos 5 following a separate export-control fight with the Trump administration. RuntimeWire reported Wednesday that Fable 5 returned globally on July 1 while Mythos 5 remained tied to government-approved cyberdefense access; our follow-up traced how quickly that redeployment faced outside testing of the model's cyber safeguards.

The private argument was always about control

The court record already showed the policy dispute in outline. In fall 2025, the Defense Department and Anthropic began discussing deployment of Claude on the Pentagon's GenAI.mil platform, according to a March 26 preliminary-injunction order in Anthropic's Northern District of California case. The Pentagon took the position that Anthropic should allow the department, contractors and subcontractors to use all versions of Claude for all lawful uses, without usage restrictions. Anthropic eventually agreed to that structure with the two exceptions at the center of the dispute: mass surveillance of Americans and lethal autonomous warfare, the court order states.

That framing matters because Amodei's dispute with Michael was never limited to whether Claude could be useful for defense work. Anthropic had already moved well past the outsider posture that many AI safety groups still occupy. The court order says U.S. intelligence and defense agencies had used Claude since November 2024 through a Palantir partnership, that the Department of War had used Claude Gov since March 2025, and that Anthropic won a two-year agreement worth up to $200 million in July 2025 to integrate and optimize AI capabilities across the department.

Anthropic's public line was that it supported lawful national-security use of Claude while refusing the two categories it viewed as unsafe or unlawful. In a Feb. 27 statement, Anthropic said the impasse followed months of negotiations and that the disputed exceptions had not affected a single government mission to Anthropic's knowledge. Anthropic also said it believed frontier models were not reliable enough for fully autonomous weapons and that mass domestic surveillance of Americans violated fundamental rights.

The Pentagon's line was different. The department argued that a private vendor could not reserve judgment over military uses once the military had decided those uses were lawful. In the court order, Judge Rita Lin summarized the government's position as a demand that the Defense Department, rather than Anthropic, decide what functions were safe for AI tools to perform.

The February deadline turned a contract dispute into a public rupture

The critical week began on Feb. 24, 2026, when Amodei and other Anthropic representatives met with Defense Secretary Pete Hegseth and Pentagon officials, according to the court order. Hegseth told Anthropic that Claude had strong capabilities, the order says, while also saying the Pentagon had other vendors that would not seek veto power over Defense Department use. By the end of that meeting, Hegseth had given Anthropic until 5 p.m. on Feb. 27 to accept all lawful uses or face a supply-chain-risk designation.

Anthropic responded publicly on Feb. 26. The court order says Amodei stated that Anthropic understood the Defense Department makes military decisions and had not objected to specific operations. He nevertheless said Anthropic could not in good conscience provide the access the Pentagon requested.

On Feb. 27, Hegseth said he was directing the department to designate Anthropic a supply chain risk. Anthropic's same-day statement called such a label unprecedented for a U.S. frontier AI developer and said the designation, if formalized, could legally extend only to Claude use as part of Defense Department contracts, rather than all commercial use by defense contractors.

The formal letter came March 4. In a March 5 statement from Amodei, Anthropic said it had received confirmation of the supply-chain-risk designation and saw no choice except to challenge it in court. Amodei also tried to narrow the operational stakes, writing that Anthropic's concern was with high-level usage categories rather than operational decision-making.

Anthropic sued on March 9. The complaint alleged that the government had punished Anthropic for its safety views and said the supply-chain-risk letter offered no specific explanation for the risk Anthropic supposedly posed. Anthropic also said federal contracts were already being canceled and private contracts were being put in doubt, jeopardizing hundreds of millions of dollars in near-term business.

The judge gave Anthropic a reprieve, then the model fight moved to export controls

On March 26, Judge Lin granted Anthropic a preliminary injunction. The order said the government could stop using Claude and seek another vendor. The legal problem, Lin wrote, was that the government went further by attempting to bar Anthropic from federal work, pressure military contractors to cut commercial ties and label Anthropic a supply-chain risk in a category historically associated with adversaries and sabotage.

The ruling did not end the standoff. It bought Anthropic room to keep selling and litigating while the Pentagon and Anthropic kept arguing over operational control, model access and the scope of safety policies.

That pattern repeated in June, when the government issued a separate directive involving Anthropic's newest models. In a June 12 statement, Anthropic said the U.S. government had ordered suspension of access to Fable 5 and Mythos 5 by any foreign national, including foreign-national Anthropic employees. Anthropic said the directive arrived at 5:21 p.m. Eastern and did not provide specific details of the national-security concern. Because Anthropic said it could not reliably verify nationality in real time, Anthropic disabled access to both models for all customers.

On June 30, Anthropic said the export controls had been lifted. In its redeployment post, Anthropic said Fable 5 would return July 1 across Claude Platform, Claude.ai, Claude Code and Claude Cowork, while Anthropic would re-enable access on AWS, Google Cloud and Microsoft Foundry as quickly as possible. Anthropic also said it would expand joint work with government partners, provide compute for government testing and make its safety and red-teaming expertise available for AI evaluation.

That concession is the practical center of the current phase. Anthropic is fighting the government's broad power to punish or restrict Anthropic while offering deeper government access to test, evaluate and shape safeguards before deployment. Amodei has also pushed in the same direction publicly. In Anthropic's Policy on the AI Exponential, Anthropic proposed that governments have authority to block or deter dangerous deployments of the most capable models, paired with transparency, independent evaluation and security requirements.

Anthropic's scale makes the fight harder to contain

Anthropic can no longer treat government access as a narrow channel. Anthropic says it raised $65 billion in Series H funding in May at a $965 billion post-money valuation and that run-rate revenue crossed $47 billion earlier that month. Those are Anthropic-reported figures, but they show why a fight over defense contracting now hits the same company that is selling Claude into enterprises, developer workflows and national-security systems.

The Pentagon also has leverage because frontier AI labs want government workloads, classified deployments and policy influence. Anthropic's own product materials now describe Claude as a family of models spanning Fable for advanced knowledge work and Mythos for cybersecurity and biology research. Anthropic says Mythos 5 is available only to a small group of vetted partners after government approval, while Fable 5 is the same underlying model with added safeguards for cybersecurity and biology.

The emails described by the Journal add texture to a record that already pointed to the same conclusion: Amodei's safety doctrine has become a commercial term that powerful buyers can accept, negotiate or reject. The Pentagon's answer has been that a vendor's red lines create operational risk. Anthropic's answer has been that some uses are beyond what current models can safely do, and that the government should test dangerous systems before they spread.

That argument will keep recurring as frontier models move from office assistants into software exploitation, intelligence analysis and autonomous workflows. The unresolved question for Anthropic is whether Amodei can keep selling Claude into the most sensitive parts of government while reserving the right to say where Claude should stop.