Anthropic is helping roll out its new Mythos model inside US government national security and cybersecurity agencies, even as the company fights the Pentagon in court over how its technology should be used. The arrangement, reported by Financial Times, shows how quickly frontier AI has moved from product demos to sensitive public-sector workflows – and how awkward that can get when a company wants both government business and strict guardrails.
According to the report, several Anthropic engineers are working directly with agencies in a hands-on support role, tailoring the model for specific tasks on site. That kind of deployment is less flashy than a launch event, but far more revealing: the real competition is no longer just about model quality, it is about who can safely embed that model inside institutions that care about access control, audit trails, and the occasional national-security headache.
Mythos is being tuned for cybersecurity work
Anthropic has not commented publicly on the details, and the government side is staying equally quiet. Still, the fact pattern is clear enough. The company is building Mythos into environments where cyber defenders want help spotting software weaknesses, triaging alerts, and making faster judgments without handing over the keys to the kingdom.
That is a crowded field. OpenAI, Google, and Microsoft have all spent the past two years pushing AI deeper into enterprise and public-sector use, while governments in the US, Europe, and Asia have been scrambling to decide what counts as useful automation and what counts as an unacceptable risk. Anthropic’s pitch is familiar: powerful model, stronger safety posture, fewer nightmares.
The Pentagon dispute is the awkward backdrop
The timing is doing a lot of work here. Anthropic is still in a legal fight with the Department of Defense after opposing use cases it considers too sensitive, including mass surveillance and autonomous weapons. The Pentagon’s earlier decision to label the company a ”supply chain risk” turned that policy disagreement into a formal dispute, which makes current cooperation with national-security agencies look less like a neat partnership and more like a carefully choreographed truce.
There is a broader precedent here: governments routinely want the capabilities of frontier AI before they fully trust its behavior, and vendors usually end up walking a line between capability and restraint. That tension is especially visible in cybersecurity, where the same model that helps identify vulnerabilities can also be asked to find them faster than defenders can patch them.
Mythos access is expanding beyond the US
Mythos is also widening its reach. Anthropic said last week that access to the model would expand from a small group of organizations in the US and the UK to 150 organizations in 15 countries. That is the clearest sign yet that the company wants Mythos positioned as a serious platform, not a niche experiment for a handful of trusted testers.
- Initial access: a limited set of organizations in the US and the UK
- Expanded access: 150 organizations in 15 countries
- Primary use case: cybersecurity and vulnerability analysis
That expansion also fits the direction of travel in Washington. Donald Trump signed an order last week creating a voluntary review process for new AI models before public release, and it also asks federal agencies to develop methods for assessing AI cyber capabilities. Translation: the state wants more visibility, more testing, and fewer surprises from systems that can write code, scan systems, and potentially break things at scale.
If Anthropic really is heading toward an IPO that could value it at more than $1tn, these government relationships stop being side quests. They become part of the story investors will price: not just how smart Mythos is, but how well the company can sell advanced AI into regulated institutions without blowing up its own safety message. The next question is whether more agencies want the model’s help badly enough to tolerate the company’s caution, or whether that caution becomes the feature they actually pay for.