Financial Times reports that Anthropic has embedded around half a dozen forward-deployed engineers within the NSA to help the agency deploy Mythos — Anthropic's flagship model — for offensive cyber operations. This isn't a research partnership or advisory arrangement. These engineers are inside the intelligence community, working directly on deployment.
The Pentagon's Legal Shadow
This deployment doesn't happen in a vacuum. Anthropically is currently locked in a legal battle with the Pentagon over Claude model export controls. The Defense Department has sought to restrict advanced AI model exports under emerging cybersecurity regulations, arguing that powerful models could enable sophisticated attacks by state actors. Anthropic has pushed back, claiming these restrictions are overly broad and would cripple legitimate AI safety research.
The embedding arrangement appears to be a compromise of sorts — Anthropic gets a foothold inside the government it needs to influence policy, while the NSA gets direct access to the model's inner workings. It's the kind of deal that would have seemed unthinkable a year ago.
The question isn't whether AI companies will work with governments. It's whether they'll do it transparently or in shadow.
What "Offensive Cyber Operations" Actually Means
Mythos deployed for offensive cyber operations likely means vulnerability discovery, penetration testing, and potentially automated exploit generation. A model like Mythos can analyze vast quantities of captured network traffic, reverse-engineer proprietary protocols, and identify attack surfaces at speeds no human team can match.
This is where the rubber meets the road for AI safety. Anthropic has built its brand on responsible AI development — their constitutional AI approach, their focus on harmlessness, their public papers on alignment. Now that same technology is being put to work finding vulnerabilities in foreign networks.
The tension is unavoidable. Either you believe AI models are too dangerous to export (the Pentagon's position) or you believe they're safe to deploy (the arrangement with the NSA). You can't have it both ways.
What This Means for the Industry
This deal signals a new phase in the AI arms race. We're past the point of theoretical safety debates and into actual operational deployment. Other labs will face similar pressures — OpenAI with Microsoft ties, Google with government contracts, Meta with whatever military applications the Pentagon finds for Llama.
The smart move for other AI labs isn't to avoid government work. It's to get in early and shape the terms. Anthropic has made their bed. The rest of the industry is about to find out whether they want to lie in it.
The big unknown is disclosure. These arrangements tend to stay classified. We only know about this because of the Pentagon's legal filings. How many other deals are happening in silence?