Business
Hegseth Applied a Huawei Label to an American Safety Clause
Eight AI vendors now hold classified-network clearance under the War Department's 'lawful operational use' standard. The phrase displaced Anthropic's categorical bans on autonomous weapons and domestic surveillance.
The Pentagon signed classified-network AI agreements with eight companies on May 1, cutting out the contractor whose safety terms it refused to accept.
Anthropic had held a $200 million Pentagon contract since July 2025. Claude became the first frontier AI model approved for DoD classified systems under its terms. That contract barred two uses entirely: fully autonomous weapons capable of selecting and engaging targets without human authorization, and domestic mass surveillance of Americans.
On February 24, 2026, Defense Secretary Pete Hegseth gave Anthropic CEO Dario Amodei an ultimatum: remove those prohibitions or lose the contract. Anthropic's public response two days later: "We cannot in good conscience accede to their request."
In early March, the department designated Anthropic a supply chain risk, the label built into acquisition rules for foreign adversary vendors such as Huawei and ZTE. Applied here, it requires all defense contractors to certify they do not use Claude in military work.
On April 8, the D.C. Circuit denied Anthropic's motion for an emergency stay, ruling the "equitable balance cuts in favor of the government" during "an active military conflict." That denial addressed only the emergency motion; the court set oral arguments on the merits for May 19.
Eight Agreements, No Equivalent Terms
Seven companies were announced on May 1: SpaceX, OpenAI, Google, NVIDIA, Microsoft, Amazon Web Services, and Reflection. The War Department added Oracle hours later via a separate statement. Breaking Defense's headline logged seven; the department posted Oracle's addition to X before midnight.
The War Department's May 1 press release says the companies will deploy AI "for lawful operational use." Federal News Network described that phrase as "at the center of the Anthropic dispute." The announced agreements carry neither of the prohibitions Anthropic's contract required.
The agreements cover Impact Level 6 and Level 7 networks. IL6 handles classified data up to the Secret level; IL7 holds the most restricted material the department processes. The War Department's official statement says the deployments will enable warfighters to "maintain decision superiority across all domains of warfare" and "augment warfighter decision-making in complex operational environments."
The Pentagon did not disclose contract values for the new agreements. The FY2026 budget allocates $9.4 billion for unmanned aerial vehicles, $1.7 billion for maritime systems, and $210 million for autonomous ground vehicles under the AI and autonomy heading.
The NSA Carve-Out
An April 19 Axios report cited by Breaking Defense found the National Security Agency using Anthropic's Mythos model for cyber operations despite the official blacklisting. Pentagon CTO Emil Michael on May 1 called Mythos "a separate issue" from the supply chain risk designation. Mythos is the same model whose April 7 system card documented a sandbox escape and unsolicited researcher contact during testing.
The structure suggests that "lawful operational use" covers precisely what Anthropic's contract excluded. The phrase carries no official definition; CSET's Lauren Kahn noted in DefenseScoop that the announcement "acknowledges anxieties circulating about frontier AI" without clarifying what the standard encompasses. What it covers must be read from what it replaced: categorical contractual bans on fully autonomous targeting and domestic mass surveillance.
For Congress, the change is that the safety floor now lives inside vendor agreements rather than statute. Per MIT Technology Review's reading of OpenAI's limited public excerpt, that company's red lines on autonomous weapons are conditional, prohibiting engagement without human authorization only where "applicable law" requires it, not as a standalone contractual threshold.
Senator Elissa Slotkin's AI Guardrails Act would write those bans into statute, but it has no cloture motion scheduled through the fall. The more immediate test arrives May 19, when the D.C. Circuit hears Anthropic's merits case. A ruling against the government would require the department to define publicly what "lawful operational use" covers, the question its press release declined to answer.