Editorial illustration for Anthropic questions Pentagon trust as NSA surveillance deemed limited
Anthropic Challenges Pentagon's AI Trust Framework
Anthropic questions Pentagon trust as NSA surveillance deemed limited
The language of limits is a careful construction. When the NSA says its surveillance of Americans is “limited,” it leans on a definition so narrow that plain English, the kind you and I use, can’t fit through the gaps. Those gaps were large enough for Senator Ron Wyden to stand on the Senate floor and warn that something was wrong, though he couldn’t say what.
Large enough for James Clapper, in 2012, to look a panel in the eye and deny that the agency collected data on Americans. That denial was a lie. And that lie broke the trust of a man named Edward Snowden, who gave us the proof we needed.
Now Anthropic, the AI company built on safety promises, is looking at the Pentagon and seeing the same pattern: a trust deficit dressed in technical jargon. They don’t believe the assurances. And a growing chorus, from security experts to the pages of *The Verge*, says you shouldn’t either.
When read with a plain English dictionary, the nature of which you and I probably have and understand, we would come away with a belief that the NSA's ability to surveil Americans was very limited, in fact to the point that they're supposed to, if they realize that they are surveilling a US person, that they're supposed to immediately stop and cry foul and erase the data and all of this other stuff. There were rumors for a while that that was not really happening and there were hints and in particular Senator Ron Wyden was very vocal about going on the floor of the Senate and saying, "Something is not right here and I can't quite tell you what," or in hearings he would ask intelligence officials, "Are you or are you not collecting mass data on Americans?" Those officials would either deflect or in some cases outright lie. I believe it was one hearing in 2012 with James Clapper, who was the Director of National Intelligence at the time, where he was asked directly on this point. And he basically said, "No, we don't collect data on Americans." That was a big part of what inspired Ed Snowden to leak the data, the reports that he leaked to Glenn Greenwald and Barton Gellman and Laura Poitras as well.
The lesson of history is not that surveillance is always abused. It is that the mechanisms we build to prevent abuse are only as strong as the people guarding them. When intelligence officials stand before Congress and offer carefully parsed denials, when programs are described in language that obscures more than it reveals, the trust that underpins democratic oversight begins to crack.
Anthropic’s reluctance to hand the Pentagon the keys to its technology is not paranoia , it is the bitter fruit of a decades-long pattern. The same system that told us the NSA’s reach was “limited” turned out to be the system that was secretly collecting records on every American phone call. The same officials who swore they weren’t spying on citizens were the ones whose lies inspired a contractor to flee with state secrets.
Today, the tools are more powerful. The incentives for secrecy are just as strong. And the public memory, as always, is short.
So when an AI company says it does not trust the Pentagon, it is not rejecting patriotism. It is remembering what the government itself has chosen to forget.
Related Reading
Browse all Anthropic news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic Trump cracks down on Anthropic after Amazon tip; staff largely foreign
Open Source
Related Topic Claude Mythos highlights EU AI safety gaps, says researcher Caroli
Policy & Regulation
From Archives Google Maps uses Gemini to answer complex queries and book tables
From Archives Google Maps adds Gemini‑powered “Ask Maps” tab that can’t be turned off by default
From Archives Claude gains shared context in Excel, PowerPoint; Microsoft adds Copilot Cowork
AI Tools & Apps
From Archives Anthropic files lawsuit against U.S. government in federal court
Policy & Regulation
Common Questions Answered
What concerns has Anthropic raised about the Pentagon's AI data pipelines?
Anthropic has flagged a potential disconnect between the defense department's confidence in AI partners and the actual data sources used to train these models. The company's engineers argue that the Pentagon may be overlooking the National Security Agency's legal limitations on domestic surveillance when developing AI technologies.
How does the NSA's surveillance policy impact AI model training according to Anthropic?
According to the article, the NSA is supposed to immediately stop and erase data if they realize they are surveilling a U.S. person, creating potential complications for AI model training. Anthropic is questioning whether these legal safeguards are being consistently and rigorously applied in the Pentagon's AI development process.
Why has Anthropic been designated as a supply-chain risk by the Pentagon?
The article suggests that Anthropic is currently entangled in a legal dispute with the Pentagon that is described as 'messy' and 'fast-moving'. This designation as a supply-chain risk appears to be part of a broader tension between the AI company and the defense department regarding data usage and surveillance concerns.
Further Reading
- Internal Pentagon memo orders military commanders to remove Anthropic's AI technology from key systems — CBS News
- Pentagon in memo to employees: You can use Anthropic tools beyond six months if needed for national security work — Times of India
- The Pentagon-Anthropic dispute reflects governance failures - with consequences that extend far beyond Washington — University of Oxford
- Statement from Dario Amodei on our discussions with the Department of War — Anthropic
