Pentagon Blacklists Anthropic for Refusal, Funds OpenAI With Same Limits

Why Did the Pentagon Blacklist an American AI Lab?

The Pentagon this year branded Anthropic, the San Francisco lab behind Claude, a supply-chain risk to national security — a label previously stuck only on foreign firms like Huawei. The fight sprang from a contract clash over how the military may deploy its models, not from any breach or planted backdoor. The dispute has cut the lab from government contracts and created hiring uncertainty that bears on international markets like Canada, the U.K., or Europe.

On February 27, 2026, President Trump ordered federal agencies to drop the lab’s technology. Defense Secretary Pete Hegseth same day said the Department of Defense would tag the company as such. On March 3, the War Department notified Anthropic by letter. Just Security and Politico reported that this mark hit an American-owned firm for the first time; Congress’s research service found no earlier Pentagon use of the provision.

The law hangs on 10 U.S.C. §3252, a 2018 procurement statute. It lets the Defense Secretary bar a named entity from a narrow set of systems if he writes that the move protects national security and that milder steps won’t work. The statute authorizes contracting limits, not sanctions. Mayer Brown’s lawyers also noted a separate letter flagged Anthropic under the Federal Acquisition Supply Chain Security Act (FASCA).

The order’s fine print limits its bite. The War Department’s notice reaches only buys under 48 C.F.R. Subpart 239.73 and any Anthropic tool that fits “covered item of supply” or sits inside a “covered system” (meaning tech for spying, crypto, command of troops, or a weapon’s guts). Payroll, finance, and logistics programs sit outside the line. Mayer Brown’s team stresses the clauses bind only work on those defense contracts; a contractor can still use Claude for ordinary business.

Hegseth drew a harder line. In a statement reported by Just Security, he said “no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic.” Anthropic countered that the statute “can only extend to the use of Claude as part of [DOD] contracts.” That gap between the law’s plain text and a blanket ban later shakes its hiring plans.

The fight traces to a July 2025 contract. The Defense Department handed $200 million apiece to Anthropic, Google, OpenAI, and xAI to speed advanced AI for national security. Claude earned clearance as the first frontier model on classified networks. Talks broke when the department demanded Anthropic drop two self-imposed limits: a ban on mass domestic spying and a ban on fully autonomous weapons. The lab refused. It told Congress today’s frontier models lack reliability for lethal autonomy and that mass surveillance of Americans violates fundamental rights. The Pentagon shot back: “The military will not allow a vendor to insert itself into the chain of command by restricting the lawful use of a critical capability.”

The statute targets firms that plant espionage tools; House oversight files cite Huawei gear found near an ICBM silo as the model threat. Slapping that label on a domestic lab that merely kept its own usage rules is the precedent Politico flags. The designation cuts Anthropic from the defense pipeline the 2025 contract opened; the award and a small State Department grant vanish against its $30 billion annual run-rate revenue reported in April 2026. Both sides agree the split was over usage terms, not enemy hacking. The legal fight and the doubt it breeds set the stage for the talent questions ahead.

Hiring Pipeline Under Stress

The Pentagon’s blacklist rattles Anthropic’s ability to lure and keep AI engineers, even as the lab posted 47 open roles in the past week, one new listing about every three and a half hours on our board. The move severs the lab from government work and, following a June 15 directive restricting foreign national access to advanced models, raises questions about visa sponsorship for roles with export-control implications, as h1bvisajobs.com notes some compute-intensive roles may carry such implications. That doubt meets a hiring machine that has sponsored about 400 H-1B visas a year, with fewer than one in 20 applications denied, the best rate among AI labs, per h1bvisajobs.com.

The label grew from a June 15 directive ordering Anthropic to block foreign nationals — even those living in the U.S. — from two advanced models, Fable 5 and Mythos 5. In that day’s Bloomberg Tech interview, a speaker said the ban hit “foreign nationals, not just other countries, but those residing in the United States.” For the lab’s many H-1B holders, the June 15 directive means the advanced models they were hired to improve can be locked behind a compliance wall.

The Visa Bottleneck

Anthropic pays all immigration fees and starts the PERM green-card process within 12 to 18 months, quicker than OpenAI’s 18-to-24-month norm. Current H-1B staff can switch to the lab cap-exempt in a 2-to-4-week premium filing. Yet h1bvisajobs.com warns the annual cap lottery still traps new applicants, and that “some compute-intensive roles may have export control implications” after the security order. With the government’s flag, those implications harden into a real constraint: why would the lab sponsor a researcher from India or China for a job they cannot fully do?

The fellowship track shows the strain. Anthropic’s Fellows Program pays $3,850 a week for four months but demands applicants already hold work rights in the U.S., U.K., or Canada. The lab sponsors no visas for it. A June 24 substack analysis found two in five of the first cohort later joined Anthropic full-time, and four in five produced public research. The fellowship does not sponsor visas, and H1B sponsorship for compute-intensive roles may carry export-control implications per h1bvisajobs.com. Applications for the 2026 cohorts are reviewed on a rolling basis, but eligible locations stay limited to those three countries where the candidate already holds work clearance.

Paychecks That Still Lead the Market

High pay hasn’t stopped the uncertainty. The fellowship’s limited eligible locations hint where talent might land if U.S. sponsorship cools. Yet the lab still pays top dollar. Its average H-1B wage sits at $232,000, the highest among AI employers. Alignment researchers command $295,000 to $425,000 base. Across 344 open roles, the median salary band reaches $405,000, with a range from $65,000 to $858,000; a recent Research Engineer posting topped $850,000.

H1bvisajobs.com now advises foreign nationals at the lab to file PERM early and build an EB-1A self-petition to skip the backlog if publications allow.

Retention Pressure

Defense tech firms told staff to drop Claude after the DoD ban last week, CNBC reported. That lost commercial business feeds anxiety inside the lab. Dario Amodei’s earlier call for the government to ban risky models now reads as irony, per the June 15 briefing. Talks with the Commerce Department are underway, but until the label lifts, the lab’s willingness to sponsor visas for foreign nationals on export-touched work may have tightened.

The hiring pipeline isn’t closed — the lab’s recent weekly postings keep coming, and its immigration team still backs both H-1B and PERM tracks. But a researcher choosing between a San Francisco offer and a London or Toronto lab now weighs a U.S. government flag against a paycheck, and the scale tilts outward.

Anthropic Fights Back in Court

On the same February 27 order, Anthropic promised to sue. “We will challenge any supply chain risk designation in court,” the company said that day. The tone mixed litigation, statutory argument, and open offers to keep serving the Pentagon under narrow limits.

The fight matters because the lab is the only American firm ever publicly named a supply-chain risk; the label previously stuck to foreign adversaries, CNBC reported March 5. A precedent letting a cabinet secretary cut a domestic AI lab from defense work over policy disagreement would reshape how every developer bargains with Washington.

Anthropic’s first statement attacked the legal basis directly. “The Secretary does not have the statutory authority to back up this statement,” it wrote, pointing to 10 USC 3252. That statute, however, bars any court or GAO bid-protest review of the agency head’s call (uscode.house.gov). The lab filed anyway. On April 8, 2026, a federal appeals court in Washington denied its stay request, leaving the designation in force during litigation, according to Vucense. Separately, a federal judge in California issued a preliminary injunction blocking parts of the designation’s operationalization, though the core label remained in force.

The company refined its argument in a March 5 post after the March 4 formal letter. Anthropic called the action “legally unsound” and warned it would “set a dangerous precedent for any American company that negotiates with the government.” The post walked through the statute’s narrow scope: 10 USC 3252 exists to protect the government, not punish a supplier, and demands the least restrictive means. The lab noted the letter reaches only Claude used directly inside Department of War contracts, not outside commercial work.

“No amount of intimidation or punishment from the Department of War will change our position on mass domestic surveillance or fully autonomous weapons.” — Anthropic, February 27 statement

The quote echoes the usage limits that broke the July 2025 contract talks. DoD wanted unfettered access for all lawful purposes. “From the very beginning, this has been about one fundamental principle: the military being able to use technology for all lawful purposes,” a DoD official told CNBC on March 5. The impasse broke talks.

Anthropic paired the lawsuit with an olive branch. In the March 5 update, it offered models to the Department of War at nominal cost with continuing engineer support “for as long as we are permitted to do so” to ease transition. The post said both sides had productive conversations serving the department within the two narrow exceptions. “Anthropic has much more in common with the Department of War than we have differences,” it wrote.

The clash drew political fire. David Sacks, White House AI and crypto czar, had accused Anthropic of “woke AI” and regulatory capture, CNBC noted. The lab ignored that line, keeping pushback on statute and precedent.

The courtroom battle will outlast the spring. Anthropic’s engineers keep shipping models while a DC circuit court weighs whether a law that bars judicial review can still be challenged. The lab’s pushback is less retreat than a wager that courts or a later administration will undo a label it calls unlawful.

Defense Buyers Pick Eight Winners

On May 1, 2026, the Pentagon announced classified AI contracts with eight firms — OpenAI, Google, Microsoft, Amazon Web Services, NVIDIA, SpaceX, Reflection AI, and Oracle. The deals replaced the July 2025 award that had made Anthropic the first AI company on Pentagon classified systems; until then, Claude was the only model in those networks. Internal numbers cited by Idlen.io put Anthropic’s lost defense pipeline at $3–5 billion over five years, a sum now flowing to the eight winners.

The contracts span Impact Level 6 (Secret) and Impact Level 7 (Top Secret/SCI) environments, delivered through the GenAI.mil portal that has grown to more than 1.3 million Defense Department users since launch. Oracle joined hours after the first seven, rounding the list with its cloud infrastructure and old defense ties. Reflection AI stands out: a startup of ex-DeepMind staff valued near $5 billion in early 2026, brought in as a counterweight to big tech. The spread signals a multi-vendor strategy breaking from Cold War single-prime awards, giving the government faster iterative deployment.

Primes already embedded in military targeting — Palantir, Anduril, and Shield AI — hold billions in separate contracts tied to revised doctrine. The eight model providers feed those primes’ systems. Displaced demand splits in two: foundation model supply now controlled by the eight, and systems integration still led by specialized defense contractors building the operational tools.

Talent follows the money and the clearance. The May 1 awards split the AI industry into two tiers by willingness to accept federal use clauses, a divide that will outlive any administration. Engineers who will build for autonomous weapons or surveillance can target the eight winners and the primes. OpenAI, which signed its Pentagon deal hours after Anthropic’s exclusion, included safeguards nearly identical to what Anthropic demanded — a detail that blunts any moral high ground for defectors. SpaceX adds hardware-adjacent AI through its space systems arm. The portal’s massive user base needs sustained engineering support, pulling cleared talent from the open market toward contract holders.

The blacklist bars every Pentagon contractor from commercial activity with Anthropic, and a six-month wind-down clock runs to August 2026 for current users. The uncertainty around visa sponsorship and lost federal pipeline may push some researchers towards those markets, where the lab’s refusal to authorize lethal autonomy and mass surveillance reads as an asset, not a liability. Domestic posting strength tempers immediate outflow.

Pentagon CTO Emil Michael, speaking on CNBC, said: “It’s irresponsible to be reliant on any one partner, and we learned that that one partner didn’t really want to work with us in the way we wanted to work with them.” The comment frames the eight-vendor spread as risk management. For talent, it means more openings at firms accepting “all lawful purposes” clauses, fewer at a lab that chose court fights over compliance.

If the Five Eyes alliance harmonizes with the U.S. position, Anthropic’s enterprise ceiling drops further; if it diverges, London and Ottawa become recruiters’ markets for safety-minded researchers.

Watch the October 2026 IPO window. Anthropic’s path to public markets now rides on whether its ethical positioning converts to European and healthcare revenue, while the eight winners absorb the defense talent that once might have flowed to a single lab.

Working in AI? Zero G Talent tracks the openings: see every open Anthropic role, browse AI jobs, the companies hiring, and the people building the field.