Anthropic’s new Mythos model is moving into U.S. national-security cyber work even as the company fights the Pentagon in court over how its AI can be used. That combination says a lot about where the money, the policy fights, and the real-world demand are heading: governments want cutting-edge models for defense, while the same vendors are trying to draw lines around the most dangerous uses.
According to Financial Times sources familiar with the project, Anthropic is helping deploy Mythos inside American government structures tied to cybersecurity. Some of its engineers are working directly with agencies in a hands-on support role, tuning the system for specific tasks rather than just handing over software and walking away. The company and the defense side have not publicly detailed the arrangement, which is usually a sign that the work is sensitive, unfinished, or both.
Mythos is being tuned for cybersecurity work
The appeal is obvious. Since launch, Mythos has been discussed for its ability to spot and analyze software vulnerabilities, exactly the kind of capability that can help defenders move faster than attackers. That is also the awkward part: the same class of model that can harden systems can be repurposed to probe them, which is why governments are moving carefully and talking a lot about safeguards.
Anthropic recently widened access to Mythos beyond its original limited release in the U.S. and U.K., with plans to make it available to 150 organizations across 15 countries. That broader rollout matters because it puts the model on more desks at the same time that rivals such as OpenAI, Google, and xAI are all racing to land similar enterprise and public-sector deals. The race is no longer just about benchmarks; it is about who gets trusted first by the people guarding critical infrastructure.
The Pentagon dispute is still hanging over the deal
The timing is messy for Anthropic. The company has been arguing for limits on certain high-risk uses of its models, including mass surveillance and autonomous weapons systems, while the Pentagon has labeled it a ”supply chain risk,” and the firm has challenged that decision in court. That kind of public disagreement does not stop governments from wanting the tech, but it does force them to negotiate harder over where the boundaries sit.
There is also a bigger policy shift under way. Donald Trump signed an executive order last week creating a voluntary review process for new AI models before public release, and directing federal agencies to build methods for assessing AI cyber capabilities. In practice, that looks like a future in which security reviews become a standard part of selling advanced AI to governments, especially as procurement teams get less interested in the flashiest demo and more interested in who can keep the system from becoming the next incident report.
Anthropic’s IPO plans add another layer of pressure
All of this is happening while Anthropic is preparing for an IPO that media reports have valued at more than $1 trillion. If that number is even directionally right, the company will be under intense pressure to prove that it can sell into government and security markets without losing control of the very capabilities that made Mythos attractive in the first place. The next question is whether this model becomes a flagship for defensive AI, or another reminder that the line between cyber defense and cyber offense is thinner than the sales pitch suggests.