The intersection of advanced AI and national security just became significantly more complicated. According to a report by the Financial Times, the National Security Agency is preparing to use Anthropic's frontier cybersecurity model, Mythos, for intelligence and offensive cyber operations — even as a federal procurement ban on Anthropic's technology remains technically in effect.
The report states that Anthropic has embedded approximately half a dozen engineers directly within the NSA to assist the agency in deploying Mythos for targeted applications. What precisely those applications are remains unclear, but the NSA's mandate includes both foreign signals intelligence collection and active offensive cyber capabilities against adversaries.
Context: The Federal Ban Paradox
The situation is made more complex by prior reporting from Axios, which in April revealed that the NSA had been accessing Mythos despite a Department of Defense-level ban on using Anthropic's services through federal procurement channels. The new Financial Times report suggests this arrangement has not only continued but deepened, with Anthropic actively supporting NSA operations on-site.
When asked about the reporting, the NSA declined to confirm or deny any details. Anthropic did not respond to a request for comment from TechCrunch.
Why Mythos Specifically?
Mythos is Anthropic's specialized cybersecurity-focused model, engineered with capabilities that go beyond general-purpose assistants. Anthropic has acknowledged publicly that it restricted access to Mythos over concerns that its ability to identify and analyze vulnerabilities could be weaponized by malicious actors. The irony of that rationale is now in sharp relief: the U.S. government's most powerful intelligence agency reportedly wants Mythos precisely because of those capabilities.
The model's potential use cases in an intelligence context include accelerating vulnerability discovery, assisting with exploitation chain analysis, and supporting the kind of complex technical reasoning that previously required teams of specialized human analysts.
Why It Matters
This development highlights a central tension in the current era of AI governance: the most powerful AI models are simultaneously dual-use tools with enormous offensive potential and commercial products subject to corporate policies and procurement regulations. When agencies with classified authority need frontier AI capabilities, standard procurement channels can become obstacles rather than guardrails.
For enterprise security leaders, the Mythos situation is also a signal about where AI-assisted cyber offense is heading. If government agencies are embedding AI engineers to deploy specialized cybersecurity models for offensive operations, the sophistication and speed of AI-assisted attacks in both nation-state and criminal contexts will only accelerate. Defensive security teams should expect AI-augmented threats to become a baseline assumption, not an edge case.
The governance questions this raises — about AI companies supporting classified operations, about federal bans that are apparently porous, and about who controls dual-use AI capabilities — have no clean answers yet. But they are becoming impossible to defer.
Source: TechCrunch, citing Financial Times