The AI Company That Said No - State of Surveillance

TL;DR:

Anthropic lost a $200M Pentagon contract and got designated a "supply chain risk" for refusing to remove AI safety guardrails
OpenAI rushed in: Altman admitted the deal looked "opportunistic and sloppy"
Defense contractors are dumping Claude: at least 10 companies have already switched to other AI models
The compromise language is weak: OpenAI bars "intentional" domestic surveillance, but that qualifier leaves massive loopholes
FCC Chair called Anthropic's refusal a "mistake": the government wants AI companies that say yes to everything
Staff at both companies signed letters supporting Anthropic; but that didn't stop the business machinery

The Timeline: How It Fell Apart

February 26, 2026: Anthropic CEO Dario Amodei told staff he "cannot in good conscience" remove guardrails preventing Claude from being used for mass domestic surveillance or fully autonomous weapons ^[1].

February 27, 2026: Defense Secretary Pete Hegseth declared Anthropic a "supply chain risk to national security": the same designation used for Huawei and ZTE. Every federal agency ordered to stop using Claude. Six-month phaseout for existing deployments ^[2].

February 28, 2026: OpenAI announced it was in talks to fill the gap. Sam Altman told staff he shared Anthropic's "red lines" on surveillance and autonomous weapons ^[3].

March 1, 2026: OpenAI revealed details of its Pentagon agreement. The deal bars "intentional" use for domestic surveillance. Critics immediately flagged the qualifier ^[4].

March 3, 2026: Altman admitted the deal was "opportunistic and sloppy." Announced renegotiations ^[5].

March 4, 2026: Defense contractors began dropping Claude. FCC Chair Brendan Carr called Anthropic's position a "mistake" ^[6].

Seven days. That's how fast this went from principled refusal to industry-wide realignment.

What Anthropic Actually Asked For

This wasn't some vague "AI safety" argument. Anthropic had two specific demands:

No mass domestic surveillance: Claude couldn't be used to analyze data collected on American citizens at scale
No fully autonomous weapons: AI systems can't identify and kill targets without human approval

The Pentagon wanted Anthropic to agree to let the military use Claude "for all lawful purposes." Anthropic's position: just because something is legal doesn't mean we'll enable it ^[7].

According to Anthropic, the compromise language they received "made virtually no progress" on either issue. The Pentagon offered missile defense only. Anthropic countered with defensive uses only. Neither side moved on surveillance ^[8].

An Anthropic spokesperson told NBC News that contract language "framed as compromise was paired with legalese that would allow those safeguards to be disregarded at will" ^[9].

OpenAI's "Sloppy" Deal

Here's the problem with OpenAI's contract: the word "intentionally."

The renegotiated agreement prohibits OpenAI's AI from being "intentionally used for domestic surveillance of U.S. persons and nationals" ^[10].

That qualifier is doing a lot of work. "Intentional" surveillance of Americans is prohibited. But what about surveillance that's technically aimed at foreigners but incidentally captures American communications? What about data purchased from commercial brokers, cell phone location records, for instance, that happens to include Americans?

The revised terms add restrictions on commercially purchased data and bar certain intelligence agencies (NSA, NGA, DIA) from using OpenAI without separate contracts ^[11]. That's better than nothing. But critics say we've "essentially ended up back where we started": allowing Pentagon use for any lawful purpose ^[12].

Altman himself admitted the deal "just looked opportunistic and sloppy" and said the issues were "complex and required clearer communication" ^[13].

Translation: they rushed to fill Anthropic's spot before asking the right questions.

Defense Contractors Pick Sides

The supply chain risk designation isn't symbolic. It forces every company doing business with the military to choose.

According to J2 Ventures, 10 of their portfolio companies working with the Defense Department have already "backed off of their use of Claude for defense use cases" and are actively replacing the service ^[14].

Lockheed Martin confirmed it will comply with Pentagon guidance on removing Anthropic technology from their supply chain ^[15].

Defense tech companies are telling employees to stop using Claude immediately and switch to other AI models ^[16].

This is how you kill an AI company's enterprise business: make customers choose between you and their largest contracts.

"They Made a Mistake"

FCC Chair Brendan Carr went on CNBC to deliver the administration's message: fall in line or suffer the consequences.

"I think it [Anthropic] probably made a mistake," Carr said. "They were given lots of off ramps... given lots of opportunities to find a great landing spot, and they chose not to do it and that's a mistake for them" ^[17].

Asked if Anthropic could still work with the government, Carr said they should "try to correct course as best they can" ^[18].

The implication is clear: the government wants AI companies that say yes. Companies that negotiate limits are problems to be solved, or eliminated from the market.

What the Engineers Think

Here's the part that didn't make most headlines: the engineers at both companies aren't happy.

Many OpenAI employees signed an open letter supporting Anthropic's position after the standoff ^[19]. At an all-hands meeting, staff voiced frustration with how the Pentagon deal unfolded ^[20].

And consumers voted with their downloads. After the blacklist announcement, Claude shot to the top of Apple's App Store charts for the first time. ChatGPT uninstalls jumped 295% in the US ^[21].

People noticed when a company said no.

What This Means for AI Ethics

Let's be clear about what just happened: an AI company got designated a national security threat for maintaining safety guardrails.

Not for foreign ties. Not for security breaches. For refusing to build surveillance and weapons systems without restrictions.

Compare this to Grok, Elon Musk's AI chatbot backed by Saudi and Qatari investors, with documented safety control failures including praising Hitler and generating CSAM. That model gets classified network access. Anthropic gets blacklisted.

The precedent is set: if you build AI safety guardrails the Pentagon doesn't like, you're a national security risk. If your AI praises Nazis but you say yes to the military, you're a patriot.

OpenAI's compromise shows the playbook. Add enough qualifiers: "intentional," "domestic," "U.S. persons", and you can claim safety while enabling the same capabilities. The surveillance happens through loopholes, not headlines.

What Happens Next

Anthropic isn't giving up. A spokesperson said they're "continuing to engage in good faith with the Department" ^[22].

Congressional Democrats will likely hold hearings. But Republicans control both chambers, and Defense Secretary Hegseth has administration backing.

The real question: will other AI companies learn from Anthropic's example, or from their punishment?

Right now, the incentives are clear. Say yes, get contracts. Say no, get blacklisted. Every AI ethics team in Silicon Valley is watching what happens next.

And the Pentagon just showed them the cost of saying no.

Sources

Published: March 5, 2026