NSA using Claude Mythos for 'offensive cyber operations,' report claims — says 'half-a-dozen' Anthropic engineers embedded inside the agency

Just months after the U.S. Department of Defense (DOD) labelled Anthropic a "supply chain risk", completely cutting off the firm from being a vendor of any AI wares, a new report claims that the National Security Agency (NSA) is using Anthropic's cybersecurity-focused Mythos model "for offensive cyber operations" with the help of "half a dozen" staff from Anthropic embedded inside the agency. The report comes from Financial Times (FT), which cites "people familiar with the arrangement."

According to the report, the rationale behind the move is that the U.S.' adversaries will surely be using cybersecurity-focused AI models like Mythos, and gaining an early upper hand is fundamental. Regardless of political optics, this logic has some merit in the face of a security landscape that's shattering the usual 90-day vulnerability report window and bringing the zero-day exploit timeframe down to a nice round zero. It's worth noting that the FT's contacts are unaware as to whether the embedded staff are assisting in any active operations (as opposed to simply guiding the use of the technology and customizing models), but they expect the bot to be useful for "infiltrating the networks of nations such as China or Iran."

The article states that Anthropic has posted a handful of "embedding engineers" inside the NSA to help with using and customizing Mythos for "specialized applications", likely meaning tailoring it to unique types of cyber-attacks. This news is surprising, as for the time being, Anthropic is theoretically banned from offering services to any branch of the DOD. It's not only DOD representatives who will be staring at their proverbial plates, either, as the cheeks of Anthropic's execs might likewise be tinted pink.

The dispute started in January 2026, a time during which the DOD and Anthropic were negotiating a $200 million contract. The Trump administration demanded that Anthropic allow usage of its wares for "all lawful purposes", implying removing some or all AI guardrails, a move that went against the firm's usage policy. This caused Anthropic to refuse to comply with the request, and CEO Dario Amodei to post a statement defending the company's position against its tech being used for tasks like mass surveillance and autonomous weapons, noting that "in a narrow set of cases, [Anthropic believes] AI can undermine, rather than defend, democratic values."

The DOD promptly shot down the contract, and eventually, in early March 2026, the Trump administration took the unprecedented step of labeling Anthropic, an American company, as a supply chain risk, putting it in the same bag as Huawei and ZTE. Anthropic proceeded to file a currently ongoing lawsuit against the DOD, calling the decision a violation of the First Amendment. The DOD went on to continue work with Anthropic rival OpenAI instead, something that may have granted a financial foothold to the revenue-embattled firm. Ironically enough, after public backlash, OpenAI added some safety clauses to its deal, which the DOD accepted, suggesting that its fracas with Anthropic may have been politically motivated.

The news also comes hot on the heels of the Trump administration issuing an executive order that asks AI companies to volunteer their leading-edge models to the government for 30 days before they're made available to other entities, though it's unclear what will be classified as a "covered frontier model" or a "trusted partner", and what happens if a company doesn't play ball.

There's no telling if the NSA purportedly using Mythos is a sign that the DOD may be warming up to Anthropic again, or if there's a schism inside the agency that believes Mythos is a first-rate necessity. The ongoing nature of the two Anthropic-DOD suits also muddies the waters even more, as two separate courts have so far issued contradictory rulings on the matter.

Mythos is currently in an early-access stage, as Anthropic announced it would distribute it to 150 organizations across 15 countries, up and out from only a handful of U.S. organizations after its April release. The news follows a revelation from Anthropic this week claiming that Claude AI is now helping to build itself at a much faster rate than expected, prompting the company to call for the option to slow down or even halt AI development.

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.