Anthropic AI Model Mythos breached classified US systems within hours

A new report from Washington has raised fresh concerns over how quickly frontier AI systems may be moving from research tools to national security flashpoints.

Senator Mark Warner, the vice chair of the US Senate Intelligence Committee, said on June 11 that General Joshua Rudd, who leads both the National Security Agency and US Cyber Command, personally told him that Anthropic’s advanced AI model Mythos had broken into most classified networks in the United States. According to Warner’s account, the breach did not take weeks. It reportedly happened within hours.

The claim arrived alongside a major shift in US policy toward Anthropic’s most powerful systems. The US government has reportedly blocked foreign users from accessing Mythos 5 and Fable 5, two of the company’s leading AI models.

The restriction appears to go far beyond ordinary commercial limits. Access was reportedly cut off not only for international public users, but also for close US allies, including members of the Five Eyes intelligence alliance: Australia, the United Kingdom, Canada and New Zealand.

That sudden decision disrupted months of work by allied governments, banks and large companies that had been trying to secure permission to use the models. The UK’s AI Safety Institute, which tests new AI systems on behalf of the British government, also reportedly lost access.

The ban does not apply to Mythos Preview, an earlier version of the system that Anthropic had already limited to a smaller group of approved users after it was assessed as a potential national security concern. Allied governments are said to still have access to that version.

A former British intelligence official suggested that access to Mythos will likely be restored for intelligence agencies, with government-level negotiations already underway. For now, however, the move shows how quickly AI access can become a matter of state control rather than product availability.

The decision has drawn comparisons with earlier US efforts to contain powerful technologies. From the 1970s through the 1990s, Washington treated public-key cryptography as a technology close to weapons-grade export control. In one well-known case, a programmer was investigated by the FBI over alleged arms trade violations before civil liberties groups helped secure broader rights to use, sell and export encryption systems.

Others have compared the current AI restrictions to the McMahon Act of 1946, which ended US nuclear cooperation with foreign governments, including the United Kingdom, despite the two countries having worked together on the atomic bomb during World War II. That cooperation was only restored after Britain developed its own nuclear arsenal.

The AI comparison is not exact, but it points to the same underlying question: how long can the United States keep its strongest technological advantages behind national borders?

The source material notes that Washington currently holds a clear lead in frontier AI, while China is believed to be roughly a year behind due to US restrictions on advanced chips. That advantage appears to be shaping the government’s response, especially as the most capable AI models become harder to separate from cybersecurity, intelligence and military planning.

The restrictions also come during a shifting period for US AI policy under President Donald Trump. In recent months, he has reportedly rolled back most of the previous administration’s AI regulations, later allowed sales of advanced AI chips to China, and on June 2 signed an executive order calling for a voluntary framework under which AI labs would give the government early access to their newest models before public release.

At the same time, the government’s Center for AI Standards and Innovation, which evaluates dangerous capabilities in advanced AI systems, has reportedly been told to stop publishing its reports. That move may be temporary, but it adds to the sense that US AI policy is being redrawn under pressure from both national security concerns and global competition.

For the tech and gaming industries, the story matters beyond Washington. Advanced AI models are increasingly tied to cloud infrastructure, cybersecurity, content creation, moderation, simulation and automation. If governments begin treating top-tier AI systems like strategic assets, access to those tools may depend less on market demand and more on security clearance, export rules and political alliances.