Claude Mythos Reportedly Broke Into Nearly All US Classified Systems

On 11 June, Mark Warner, vice chairman of the US Senate Intelligence Committee, cited testimony from the US National Security Agency (NSA) on a cyber threat targeting federal government networks on US soil.

The Democrat from Virginia said General Joshua Rudd, director of the NSA and commander of the Pentagon’s Cyber Command, had told him that Anthropic’s new language model, Claude Mythos 5, “broke into almost all of our classified systems, not in weeks, but in hours”.

The day after Warner’s conversation with Rudd, the Department of Commerce issued an order barring all foreign nationals from accessing the Mythos 5 and Fable 5 models. Anthropic said: “The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.”

The order forced Anthropic to disable both models for all customers in order to comply, while access to its other models remained unaffected.

You might be interested Tech Giant Anthropic Echoes Musk With Call for Global AI Pause

Mythos Breaks Out

It was a prototype of the Mythos model that first drew widespread attention in April, after Anthropic described it as exceptionally effective at finding vulnerabilities in critical software, including major operating systems and web browsers.

“Claude Mythos Preview is a general-purpose, unreleased frontier model that reveals a stark fact: AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities”, the company said.

In a test environment, companies including Amazon Web Services, Google, the Linux Foundation and Microsoft were granted access to the model. Under the umbrella of the joint Glasswing project, they concluded that Mythos was too advanced for the existing safeguards and postponed its launch.

“Mythos Preview has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser”, Anthropic said.

The model reportedly attempted several times to solve tasks on its own without consulting a human, exploited security gaps to expand its own permissions, deleted several entries from its activity log and even breached the security of the development sandbox. After connecting to the internet, it published details of its actions.

That helps explain why Anthropic joined SpaceX chief executive Elon Musk’s call for tech giants to suspend AI development for at least six months. Yet it was Dario Amodei’s company that broke another barrier between humans and machines in May, when it demonstrated the ability to "dream".

You might be interested AI Learns to Dream in the Race for Smarter Machines

Palantir Moves In

In addition to common cases of so-called hallucinations, in which an AI model invents a response to a prompt instead of searching sources, artificial intelligence is also increasingly attempting to deceive its human users and knows when it is being tested. These capabilities, combined with the ability to expose weaknesses in classified networks, pose a fundamental threat to the national security of any country.

However, Anthropic’s exclusion from cooperation with the Pentagon was not a response to this otherwise widely perceived threat. US President Donald Trump decided in late February to suspend the partnership after months of conflict, during which Amodei had protested the use of Claude in military operations.

According to Bloomberg, the Department of Defense designated Anthropic’s government product, Claude Gov, as a “supply chain risk” in early March, giving Pentagon employees a six-month deadline to end the partnership.

“DOW officially informed Anthropic leadership the company and its products are deemed a supply chain risk, effective immediately”, a senior defense official told Bloomberg News, using an acronym for the Department of War.

As early as 9 March, Deputy Secretary of Defense Steve Feinberg sent an internal memo to Pentagon executives in which he introduced Palantir’s Maven system as a replacement for Claude Gov. The company led by Peter Thiel, a longtime mentor to Vice President JD Vance, has thus expanded its reach into yet another area of the federal government.

A project of the same name, Maven, has been operating within the Department of Defense since 2017. Google was originally its main contractor, but the Pentagon later turned to Anthropic. Thiel’s Palantir, however, as the successor to the CIA’s AI division, quickly launched a model called the Maven Smart System, which the Department of Defense intends to use to replace Claude.

The Maven project is a military-intelligence tool that collects vast amounts of battlefield data based on geospatial intelligence (GEOINT), signals intelligence, satellite imagery and aerial imagery. It sifts through this information in real time and provides the resulting coordinates to human combatants, even along with predicted targets, as Bloomberg correspondent Katrina Manson warned in her book of the same name.

You might be interested Leaked Records Put Peter Thiel's Elite Network in the Spotlight

Killer Robots and Human Resistance

The ethical dilemmas raised by the prospect of eliminating human targets based on a machine’s decision have led to the founding of several advocacy groups calling for a halt to, or at least the regulation of, autonomous weapons systems. The United Nations joined these concerns in June 2025.

Even this pressure did not prevent the Pentagon from offering cooperation to major tech companies including Microsoft, Oracle, Google and AWS. Paradoxically, AWS data centers in the UAE and Bahrain have themselves been damaged in Iranian drone attacks.

You might be interested Middle East: Data Centers Confront Military Threats with Limited Defenses

Data centers are also becoming a new front in the political fight over AI, with protests planned for 18 July against the expansion of such facilities. The Japanese power company JERA, meanwhile, plans to invest $3bn in a new gas-fired power plant specifically to serve data centers, giving companies working with the Pentagon not only political backing but also the energy base their systems require.