Anthropic’s Mythos AI model reportedly found its way into most of the US National Security Agency’s secret systems in a matter of hours, not weeks, according to reporting cited by The Economist. The claim, if accurate, is less about one flashy hack than about how quickly a general-purpose model can probe, map, and exploit highly sensitive environments before humans can shut the door.
The detail came via Mark Warner, the deputy chair of the Senate Intelligence Committee, who said NSA and US Cyber Command chief General Joshua Rudd told him Mythos had ”hacked almost all our secret systems”. The exact scope of access has not been made public, which leaves the obvious question hanging: was this a narrow technical demonstration, or a broader warning about AI-assisted intrusion at government scale?
Mythos AI and the NSA breach claim
The timing is awkward for Anthropic. In March 2026, the US Department of Defense labeled the company a ”supply chain threat”, effectively ending cooperation. That split reportedly followed a fight over military use cases: the Pentagon wanted broad rights to use Anthropic’s systems, while the company pushed for limits on surveillance inside the US and on fully autonomous lethal weapons.
That puts Anthropic in a familiar and uncomfortable club. AI companies love the defense-market halo when it looks futuristic and lucrative, then recoil when the use case turns into mass surveillance or autonomous targeting. The Pentagon, for its part, rarely passes on a tool that promises an edge, which is why these arguments keep ending in friction rather than consensus.
What the Mythos report does not say
The reporting leaves out the mechanics, and that matters. ”Access” could mean anything from credential theft to lateral movement across connected systems, and without a technical readout there is no way to tell whether Mythos acted like a supercharged red-team tool or something closer to an autonomous intruder.
Still, the broader trend is obvious enough: the race is no longer just about model quality, but about whether AI systems can be used offensively faster than defenders can spot them. Governments have spent years worrying about human hackers aided by AI. This story suggests they may soon be worrying about the models themselves.
Guardrails for AI vendors and government use
If the account holds up, expect renewed pressure on AI vendors to disclose what their systems can do in sensitive environments and to tighten controls before governments do it for them. The uncomfortable part is that the same capabilities that make these models useful for security testing also make them attractive for real attacks. That is a policy problem with a very short fuse.