According to a report from Politico, the Pentagon is seeking to adopt and weaponize frontier AI models with hacking capabilities like Claude Mythos Preview.
Politico’s reporting comes from a leaked email and two anonymous sources.
The Pentagon is theoretically not supposed to use Claude Mythos Preview openly. While Anthropic’s still-unreleased model is ostensibly too scarily powerful to release to the public—which surely makes it extra intriguing to the military—it’s also controlled by a company the Pentagon melodramatically shunned earlier this year. Though litigation is still pending, the Pentagon has officially designated Anthropic a supply chain risk after a prolonged public disagreement, and—officially at least—must work with its competitors instead. In theory, the designation means not only isn’t the Pentagon allowed to work with Anthropic, but neither is any Pentagon contractor.
Politico says this new AI initiative within U.S. was announced earlier this month by Joshua Rudd, who is Cyber Command’s leader, and also the director of the NSA. Those two agencies are both apparently involved with this task force, and their aims sound a bit like Anthropic’s Project Glasswing, which allows certain companies and organizations to use Claude Mythos Preview to dig for vulnerabilities in their systems, and help patch those vulnerabilities.
Politico says the task force will “study how the Pentagon can safely deploy leading AI models in all aspects of its missions,” and assess “how AI models built by Silicon Valley tech giants can be used on ‘high-side’ systems bearing some of the intelligence community’s most sensitive secrets.”
But it also sounds from the report like the Pentagon task force has an additional dimension that’s not currently part of Glasswing: weaponizing AI models with hacking capabilities. A former Cyber Command deputy commander, Charles Moore, told Politico, “AI tools are rapidly becoming essential for detecting threats, prioritizing vulnerabilities, accelerating decision making and conducting both defensive and offensive cyber operations faster than our adversaries.”
Politico also says Anthropic has told the publication in the past that it is open to having Mythos used in offensive operations.
Despite its feud with Anthropic, the Trump Administration has some involvement with Project Glasswing. A Wall Street Journal article from late last month says the White House is “involved in the rollout of Mythos because of the national-security risks posed by the model,” and that it for some reason had veto power over Anthropic’s attempt to add 70 new companies and organizations to Project Glasswing.
According to Politico, use of Mythos is “something the Pentagon is likely to explore.”
