- The Anthropic NSA partnership puts Mythos AI at the center of offensive cyber operations targeting China and Iran.
- Anthropic NSA engineers — roughly half a dozen — are embedded directly at the agency to adapt and support the model.
- Anthropic is simultaneously fighting the Pentagon after the DoD labeled it a supply chain risk over AI usage limits.
- The NSA contract was exempt from the Pentagon ban, highlighting deep tensions inside US government AI procurement.
- The Anthropic NSA partnership puts Mythos AI at the center of offensive cyber operations targeting China and Iran.
- Anthropic NSA engineers — roughly half a dozen — are embedded directly at the agency to adapt and support the model.
- Anthropic is simultaneously fighting the Pentagon after the DoD labeled it a supply chain risk over AI usage limits.
- The NSA contract was exempt from the Pentagon ban, highlighting deep tensions inside US government AI procurement.
Anthropic NSA: The Mythos Model Goes to War
The Anthropic NSA relationship just got a lot more concrete. According to a Financial Times report, the National Security Agency is actively using Anthropic’s Mythos AI model to conduct offensive cyber operations — operations that reportedly include targeting networks in China and Iran. This isn’t a pilot program or a theoretical proof of concept. It’s live, it’s operational, and Anthropic has engineers sitting inside the NSA to make it work.
Mythos is a separate model from Anthropic’s publicly available Claude lineup. It’s been quietly expanding — Anthropic recently opened access to around 150 organizations across 15 countries — but its role at the NSA represents the most operationally significant deployment of the model that’s been publicly confirmed. What exactly it’s doing inside those operations remains deliberately vague. The FT report stops short of clarifying whether Anthropic’s embedded engineers are directly involved in active offensive missions or whether their role is purely technical support and model adaptation. That’s a distinction that matters enormously, both ethically and legally.
Engineers Inside Fort Meade
Anthropic has placed approximately six engineers directly at the NSA. That’s not a liaison arrangement or a quarterly check-in — that’s sustained, on-site integration. When an AI company embeds its own staff inside one of the world’s most secretive intelligence agencies, it signals a level of operational dependency that goes well beyond a typical software contract. The NSA needs people who understand the model’s architecture, its failure modes, its edge cases. And Anthropic, presumably, needs visibility into how its technology is actually being used.
The Anthropic NSA arrangement also raises questions that the company hasn’t publicly addressed. What guardrails, if any, apply to Mythos in this context? Claude, Anthropic’s flagship public model, comes with detailed usage policies — restrictions around harmful content, surveillance, autonomous weapons. Does Mythos operate under similar constraints when it’s being used to break into foreign government networks? The company hasn’t said.
A Legal Fight With the Pentagon Runs in Parallel
Here’s where things get genuinely complicated. At the same time Anthropic is embedded with the NSA, it’s in an active legal dispute with the Department of Defense. The DoD classified Anthropic as a “supply chain risk” — a designation typically reserved for foreign adversaries or companies with known security vulnerabilities — and tried to cut the company from federal contracts entirely. The reason? Anthropic had pushed back on allowing its Claude models to be used for mass surveillance programs and autonomous drone targeting.
That’s not a trivial disagreement. It reflects a fundamental tension between what AI companies say their values are and what the US military-intelligence complex actually wants from them. Anthropic drew a line around two specific use cases: population-scale surveillance and lethal autonomous systems. The Pentagon, apparently, didn’t want those lines drawn for them.
The NSA’s Mythos contract was specifically exempt from that ban, which tells you something about how these procurement fights actually work. Different agencies, different contracts, different exemptions. The DoD’s attempt to blacklist Anthropic didn’t reach the NSA’s existing arrangement. Whether that exemption was negotiated deliberately or simply fell outside the DoD’s jurisdiction isn’t clear — but the result is that Anthropic finds itself simultaneously suing one part of the US government while having its engineers embedded in another.
Does This Conflict With Anthropic’s Stated Values?
Anthropic would probably say no — and there’s a defensible logic to that position. The company’s public objections to the Pentagon haven’t been about refusing all national security work. They’ve been specifically about protecting US citizens from AI-enabled surveillance and about preventing fully autonomous lethal systems. Offensive cyber operations against China and Iran don’t obviously fall into either of those categories.
The Anthropic NSA collaboration, from the company’s perspective, might look entirely consistent with its stated mission. Claude’s constitution — Anthropic’s public framework for model behavior — frames the company’s goal as ensuring AI benefits humanity. Helping the US government conduct cyber operations against geopolitical adversaries could reasonably be framed, internally at least, as contributing to national security in ways the company considers legitimate.
But that framing deserves scrutiny. Offensive cyber operations aren’t a clean category. They can mean everything from network reconnaissance and intelligence gathering to destructive attacks on critical infrastructure. The line between those things is political as much as technical, and it shifts depending on who’s drawing it. An AI model capable of autonomously identifying and exploiting vulnerabilities in foreign networks is a significant capability — one that could be used proportionately or not, depending entirely on the humans giving the orders.
The Bigger Picture: AI and the Intelligence Industrial Complex
Anthropic isn’t alone in navigating this terrain. Google’s work with the US military — Project Maven, which used AI to analyze drone footage — triggered a staff revolt in 2018 and eventually led the company to decline a contract renewal. Microsoft has been far less conflicted, signing major cloud and AI contracts with the Pentagon and intelligence agencies with relatively little internal friction. OpenAI, after initially barring military use of its models, quietly updated its usage policies in early 2024 to allow national security applications.
The pattern is clear: the commercial AI industry is being pulled steadily into the US national security apparatus, and the companies that initially tried to maintain distance are finding that position harder to hold. For Anthropic, the Anthropic NSA relationship represents a meaningful step in that direction — more significant than a cloud hosting deal, more operationally intimate than a research partnership.
President Trump’s executive order on voluntary AI safety testing, which Anthropic publicly welcomed, adds another layer of context. The order signals that the current administration wants AI companies as partners in national security, not skeptics standing at arm’s length. Anthropic’s positive response to that order, combined with its NSA work, suggests the company is making a strategic choice to engage rather than resist.
What Comes Next
The legal fight with the Pentagon will eventually resolve — either through the courts, a negotiated settlement, or a change in DoD policy. But the more lasting question is what it means for AI safety culture when the companies most vocal about responsible AI development are also embedding engineers inside offensive cyber programs.
Anthropic has built its brand on being the safety-first AI lab. That reputation has real value — it attracts talent, shapes regulation, and differentiates the company from competitors who’ve been less vocal about risks. Whether that reputation can coexist with deep integration into the NSA’s offensive operations is something the company will have to answer, sooner or later, in public. Right now, the Anthropic NSA story is mostly being told in fragments — a Financial Times report here, a legal filing there. But as AI becomes more central to signals intelligence and cyber warfare, the pressure for transparency is only going to increase.
