US government agencies have been using Anthropic's artificial intelligence (AI) model "Mithos" for several months, it has been confirmed. Despite President Donald Trump's directive telling all government agencies not to use Anthropic products, and the US Department of Defense designating Anthropic as a supply chain risk company, the agencies chose Anthropic again. The reason is the outstanding performance of Anthropic's AI model. The US government is reported to have used Mithos to identify numerous vulnerabilities in government software.
Cybersecurity Agency Follows NSA: "We Use Mithos"
Reuters reported Thursday, citing multiple sources, that the Cybersecurity and Infrastructure Security Agency (CISA), under the US Department of Homeland Security, has been using Mithos to audit government software.
According to sources, CISA is monitoring government code repositories with Mithos to find security vulnerabilities that foreign spies and cybercriminals could exploit. Reuters said it could not confirm exactly how much government code this evaluation team has audited, but that numerous vulnerabilities were discovered, according to multiple sources.
The White House and Anthropic have been in conflict since early this year over authority regarding the scope of AI use. The dispute originated from the US military's use of Anthropic's Claude during the operation to arrest former Venezuelan President Nicolás Maduro. At the time, Anthropic argued that its AI model should not be used for large-scale surveillance or autonomous lethal weapons, but the Department of Defense countered that it should be usable for all legal activities. As the conflict between the two sides reached its worst point, the US Department of Defense designated Anthropic as a supply chain risk company. The supply chain risk company classification had previously been applied only to foreign companies accused of facilitating espionage activities.
However, the tension appears to have eased somewhat after the US District Court for the Northern District of California blocked such measures and Anthropic released Mithos, which excels at finding cybersecurity vulnerabilities. Earlier, media reports indicated that the US National Security Agency (NSA) had also been using Mithos since April. Late last month, The New York Times (NYT) reported that NSA analysts had been testing Mithos in a classified environment and were deeply impressed by its performance.
"Claude Can Have 'Separate Thoughts' It Doesn't Share With Users"
Although the Trump administration is at odds with Anthropic, an ironic situation continues in which the outstanding performance of Anthropic's AI models draws all the more attention. Earlier, by ordering export bans and release delays for Anthropic's latest AI models Mithos and Fable, the US government imprinted on the public an image of "dangerously capable models."
Amid this, Anthropic released research findings Thursday saying it had identified a small internal space where its flagship AI model Claude can think separately without expressing it to users in text.
The company explained, "Just as a human can perform one action while thinking about something else, Claude can also activate concepts and computations unrelated to its output in that space." Anthropic said this structure bears an intriguing similarity to how humans consciously access their thoughts. The company also named this area "J-space," after the Jacobian, the mathematical technique used to detect this phenomenon.
Anthropic said that observing what happens in this J-space could be a key to detecting when AI models act independently, deviating from human intent. "We can find out what Claude is thinking but not telling us," the company said. "Some of what was discovered in that process is concerning."
This research could, in the long term, connect to the question of "whether AI consciousness can be recognized." Axios noted that "Anthropic has not proven that Claude has emotions or experiences," but pointed out that "the word 'conscious' appears more than 200 times in this research."
