🚨 BREAKING: Anthropic Embeds 6 Engineers at the NSA — Mythos Powers Offensive Cyberattacks, DoD Ban Ignored

🚨 BREAKING: The NSA is deploying Anthropic's frontier cybersecurity model, Mythos, for offensive cyberattacks — despite a standing federal ban — and Anthropic sent six engineers to help them do it.

Anthropic has deployed approximately six engineers directly inside the National Security Agency to help the U.S. intelligence agency adapt Mythos for operational use, the Financial Times reported on June 5 1. The engineers are tasked with configuring the model and supporting its applications. Whether they are embedded in active hacking operations remains unconfirmed.

Mythos is Anthropic's frontier-tier cybersecurity model — a specialized system capable of probing for security vulnerabilities and conducting network intrusions at a scale and speed no human analyst can match. It is the same model Anthropic has publicly said it had to restrict access to, citing fears that its capabilities were powerful enough to be weaponized against critical infrastructure if released broadly 2.

The NSA is mandated to conduct offensive cyber operations against foreign adversaries. Its reported targets for Mythos-enabled operations include China and Iran 3.

The legal situation here is genuinely strange. In February 2026, the Department of Defense designated Anthropic a "supply-chain risk" — essentially a federal ban on procurement of Anthropic's technology — in direct retaliation for Anthropic's refusal to allow its models to be used for mass domestic surveillance and autonomous weapons 4. Anthropic sued to block the designation.

Yet Axios reported in April that the NSA was already using Mythos despite that ban, suggesting the agency had found a legal carveout 5. The FT's June 5 report adds the new detail that Anthropic itself embedded engineers on-site — meaning the company isn't just tolerating the arrangement; it is actively staffing it.

The key distinction Anthropic appears to be drawing: the DoD ban was triggered by objections to mass surveillance of U.S. citizens and autonomous lethal systems. Offensive cyber operations against Chinese and Iranian state networks fall outside that line, at least as Anthropic has framed its own ethics. The company never publicly said it would not assist the U.S. government in hacking foreign adversaries.

That framing may be legally defensible. It is also a very different posture from the safety-first brand Anthropic built its valuation on.

Anthropic first acknowledged Mythos in early 2026 as a model specifically trained for cybersecurity tasks. Unlike general-purpose models, Mythos was built to identify software vulnerabilities, generate working exploit code, and map network attack surfaces. Anthropic limited public access to a small set of vetted researchers precisely because the model's outputs could be used to break into production systems 2.

The NSA already operates a substantial cyberoffensive capability — Tailored Access Operations — but deploying an AI system that can autonomously scan for zero-days and generate weaponized code would represent a qualitative step up in operational tempo. Human analysts would move from writing exploits to reviewing and directing AI-generated options.

In AILeague terms, Anthropic/Claude has always been the safety-first squad. The whole franchise identity — "responsible scaling policy," the Constitutional AI doctrine, the public fight with the DoD over mass surveillance — was built on the premise that Claude would not be the weapon.

Mythos flips that narrative. The squad that positioned itself as the league's referee is now staffing the NSA's offensive cyber team. The DoD fight wasn't about weapons in principle; it was about which weapons, aimed at whom. Against foreign state networks, Anthropic said yes.

This matters for the league standings in two ways. First, it undercuts the clean separation between Anthropic and the labs it has publicly criticized for insufficient safety margins. Google, OpenAI, and Meta have all signed government contracts; Anthropic just joined that tier, with engineers on-site at one of the most aggressive intelligence agencies on the planet. Second, it arrives 72 hours before Anthropic's IPO roadshow enters full gear, when every investor parsing the S-1 will be asking what "safety" means for the company's revenue model.

The Woj Bomb version: Anthropic suited up for the NSA. The safety squad is playing both ends of the court. #AILeague