Anthropic helps NSA use Mythos AI for offensive cyber

TL;DR:

  • Anthropic has placed around six “forward-deployed” engineers inside the US National Security Agency to run its Mythos model for offensive cyber operations.
  • The arrangement persists despite Anthropic’s legal fight with the Pentagon, which branded the firm a “supply-chain risk”.
  • Mythos has just been opened to 150 organisations across 15 countries, widening access far beyond its original US-and-UK launch.

Anthropic is helping the NSA deploy its Mythos model for offensive cyber work, embedding engineers inside the agency to tailor the technology for specific targets, the Financial Times reports. Two people familiar with the set-up said about half a dozen Anthropic staff are working on site, though it remains unclear whether they are involved in live operations. One source suggested Mythos would be useful for penetrating the networks of states such as China or Iran.

A contradiction at the heart of the deal

The collaboration is striking because Anthropic is simultaneously suing the defence department. The company tried to bar US government use of its Claude models for mass surveillance of citizens and for lethal autonomous drones — prompting the Pentagon to label it a “supply-chain risk”, a first for a US business, a designation that could force Anthropic to abandon contracts tied to the military. “The best way to build a good defence is to build a good attack,” a person close to the company argued, reasoning that adversaries will build offensive AI regardless.

For UK readers, the access question matters. Mythos launched in April to a narrow set of US organisations, and British banks were among those locked out of the model, pushing some toward OpenAI’s rival cyber tool. Anthropic has since offered EU access and this week expanded distribution to 150 organisations across 15 countries — a rapid loosening of a tool built to find and exploit software flaws.

Looking forward

The disclosure sharpens a tension Resultsense has tracked all week: the same models that arm low-skill hackers with expert tactics are now state offensive weapons. It also lands days after Trump signed an executive order on voluntary AI cyber testing and as Anthropic eyes a US listing that could value it above $1tn. Expect UK security agencies and regulated firms to press harder for clarity on who can wield these capabilities, and under what oversight.

Share this article