The Financial Times, citing sources familiar with the matter, reported that Anthropic has deployed approximately six engineers to the U.S. National Security Agency to assist in the deployment of its unreleased Mythos model for offensive cyber operations. The report noted that the model could potentially be used for cyber infiltration missions targeting countries such as China and Iran, but whether it is directly involved in specific operations remains unconfirmed.
Mythos is only available to limited partners.
Anthropic had not previously released Mythos to the public, citing concerns about model misuse. The company granted access only to vetted partners through a restricted collaboration program called Project Glasswing. The report mentions that Microsoft, Apple, and Amazon are among those partners.
The report also stated that Anthropic has a dispute with the U.S. Department of Defense. In late February of this year, U.S. Defense Secretary Pete Hegseth listed Anthropic as a "supply chain risk," leading to the loss of a $200 million contract. The disagreement stemmed from Anthropic's refusal to allow the Department of Defense to use Claude for fully autonomous weapons or large-scale domestic surveillance, a restriction not applicable to the NSA contract.
The blacklist dispute is still ongoing.
A California judge blocked the finding, arguing it could involve retaliation against corporate statements. Subsequently, the District of Columbia Court of Appeals did not grant Anthropic's request to suspend the finding during the litigation. However, the Financial Times reported that the NSA continued to use Mythos during this period.
This news comes as Anthropic is moving forward with its initial public offering (IPO). The original report stated that the IPO could value the company at over $1 trillion.
The company also warned on the same day that AI is self-accelerating its research and development.
Almost on the same day, Anthropic's internal research institute published "When AI Builds Itself," discussing Claude's progress in automating its own development. The article stated that Claude has now written over 80% of the merged code in the company's production codebase, a figure that would be much lower before the launch of Claude Code in early 2025.
Anthropic researchers believe that AI is moving towards "recursive self-improvement," meaning that the system can not only perform tasks but also autonomously design experiments, build and train the next generation of models, and the role of humans in the development process will continue to decline.
Test results drive calls for a halt
The report mentioned that Anthropic conducted an internal test in April: the Claude agent was given an AI safety research problem, namely whether a weaker model could reliably supervise a stronger model. Two human researchers made up for a 23% performance gap in about a week; the Claude agent made up for 97% with a cumulative computing power of over 800 hours and designed all the experiments itself.
Anthropic therefore advocates for the establishment of a verifiable global AI pause mechanism, in which multiple cutting-edge laboratories simultaneously suspend operations, and each party independently verifies that they have indeed ceased progress. The company stated that it would be willing to join such a coordinated mechanism if one were to emerge.
