One of the most contentious US agencies is said to be conducting taste tests of Anthropic's highly advanced Mythos AI.

      The agency's reported use of Mythos indicates a growing divide within the US government regarding AI risks.

      The AI struggle for the US government has become increasingly complicated. The National Security Agency is reportedly utilizing Anthropic's Mythos Preview while senior officials at the Pentagon continue to advocate for severing ties with the company over supply chain issues. This highlights how real security requirements can quickly exceed official protocols.

      Since February, the Defense Department has sought to block Anthropic and encouraged other vendors to follow suit. However, as per an Axios report, the NSA seems to be proceeding with one of the company’s most advanced models, indicating that the demand for cybersecurity is taking precedence over the ongoing disputes within the government.

      Why access to Mythos is so restricted

      Mythos is notable because Anthropic is apparently keeping it under strict control. Sources indicate that the company has limited its access to about 40 organizations due to the model's offensive capabilities, with only a few users having been publicly disclosed. One source mentioned that the NSA is among those unnamed agencies granted access.

      This situation appears more like a high-stakes security tool rather than a conventional chatbot deployment. Sources have indicated that organizations with access primarily use Mythos to scan their own systems for vulnerabilities, which clarifies why national security officials remain interested despite the trust issues.

      The contradiction in Washington regarding AI

      A more significant concern lies in the evident contradiction. One segment of the government views Anthropic as a risk, while another is reportedly conducting tests of its leading model. This situation suggests that the blacklist may not be as definitive as it seems.

      The dispute seems to extend beyond procurement issues. Defense officials desired Anthropic to allow Claude to be used for all lawful functions, while the company opposed applications related to mass domestic surveillance and autonomous weaponry. Some officials interpreted this resistance as evidence that Anthropic could not be relied on in military situations, a claim the company contests.

      What comes next

      The next question is whether Mythos will remain an exception for the NSA or lead to broader acceptance within the government. Reports indicate that a recent discussion involving Anthropic CEO Dario Amodei, White House chief of staff Susie Wiles, and Treasury Secretary Scott Bessent concentrated on the use of Mythos in government and the company’s overall security stance. Both parties characterized the meeting as fruitful.

      If more agencies adopt this approach, this scenario might serve as a preview of how Washington navigates powerful AI when internal policy disputes clash with essential tools that officials are reluctant to forfeit.