Mythic Preview: Anthropic allegedly helps the NSA with offensive AI use
heise cites a Financial Times report that Anthropic staff are helping the NSA use Claude Mythic Preview for offensive cyber work.
Intelligence analysis by GPT-5.4 Mini
The story says several Anthropic employees are reportedly supporting the NSA on site as it adapts Claude Mythic Preview for offensive cyber operations. It lands amid an ongoing dispute between Anthropic and the Pentagon, even as the company argues adversaries would use similar methods.
Anthropic made a very powerful AI tool, and a report says some of its workers are helping the U.S. spy agency use it for cyber attacks. It is like teaching a super-smart lock picker how to spot weak doors, so it can also be used to break in.
Analysis
What the report says
heise reports, citing the Financial Times and two anonymous sources, that several Anthropic employees are helping the NSA use Claude Mythic Preview for offensive cyber activity. The article says the exact scope of that help is unclear, including whether the staff are directly supporting offensive operations or only helping adapt the model for specific uses.
Why the model matters
Anthropic introduced Mythic in early April and described it as so dangerous that it would only be offered to companies working in IT security. According to the article, the model is especially good at finding vulnerabilities and identifying exploitable attack paths. Anthropic later said it would broaden access to more companies.
The political backdrop
The report comes against a tense backdrop between Anthropic and the Pentagon. The article says the U.S. Defense Department had already ordered Anthropic technology removed from its own systems after the company set red lines around certain uses, and Anthropic is reportedly challenging that move in court. The reported NSA cooperation suggests the disagreement may not be as deep as it first appeared.
The company’s logic
The article says Anthropic justifies the cooperation by arguing that the company’s opponents would certainly use similar methods. One source is quoted saying the best defense is to understand how to attack, but the reporting does not confirm how much of that philosophy is actually being applied in practice.
Key points
- heise says the Financial Times reported Anthropic staff are helping the NSA use Claude Mythic Preview for offensive cyber work.
- The exact scope of the help is unclear, including whether it reaches direct operational support.
- Anthropic introduced Mythic as a highly sensitive model meant for security-focused companies.
- The article says Anthropic and the Pentagon have been in conflict over AI use, with the Defense Department removing Anthropic tech from its systems.
- The reporting suggests the disagreement may be less severe than it first looked.
If the cooperation stays limited to controlled security work, the model could help the NSA find weak spots before enemies do. That could improve defenses against hacking in a faster and more precise way than older tools.
The same capabilities could also make offensive cyber operations easier to scale, which raises risk if boundaries are vague. The reported split with the Pentagon also shows how quickly AI policy disputes can become legal and political fights.
