NSA’s Access to Anthropic’s Advanced AI Model Halted Amid Government Dispute

The National Security Agency has lost access to a high‑performance AI model created by Anthropic after a dispute with the administration, U.S. officials said. The loss deprives analysts of a tool that has impressed them with its ability to locate software vulnerabilities.

Earlier this month the administration imposed export controls on Anthropic, citing national‑security concerns. The measures forced the company to suspend the rollout of its most advanced models, dubbed Mythos 5 and Fable 5.

NSA cybersecurity analysts had been evaluating Anthropic’s tools when the newest models were disconnected.

Even within the agency’s highly secure environment, the controlled tests showed notable capabilities. The NSA’s headquarters, a clandestine facility outside Washington, focuses on developing digital‑espionage techniques and safeguarding U.S. networks from cyber threats.

The significance of Anthropic’s technology was highlighted during a congressional hearing this month, underscoring the administration’s growing reliance on cutting‑edge AI for cybersecurity while simultaneously confronting a leading U.S. developer.

Senator Mark Warner, the ranking Democrat on the Intelligence Committee, told the hearing that NSA Director Gen. Joshua Rudd had reported Mythos “broke into almost all of our classified systems, not in weeks, but in hours.” The claim attracted attention after The Economist referenced it in a report.

Officials clarified that Warner’s statement oversimplified technical details and sparked speculation on social media that the newest AI models could dramatically compromise even the most secure classified networks.

In reality, the tests were conducted by “red teams” of NSA analysts working in a tightly controlled environment that would be extremely difficult for an adversary to replicate. These teams operated within classified NSA systems that are isolated from the public internet and accessible only from specific computers.

The red‑team exercises showed that Mythos could quickly identify cybersecurity flaws within that isolated network, but it did not actually breach the systems.

Red‑team testing is a standard practice in cybersecurity, used by both private companies and government agencies to identify and remediate vulnerabilities.

Even though the feared “doomsday” scenario did not materialize, NSA analysts were struck by Mythos’s performance in the controlled tests, which exceeded already high expectations.

A White House official, speaking anonymously, said the administration is taking steps to protect classified systems from cyber threats and continues to employ advanced AI models to mitigate vulnerabilities, though the specific technology was not disclosed.

On Monday, cybersecurity agencies from the United States, the United Kingdom, Canada, Australia and New Zealand—collectively known as the Five Eyes—issued an unusual public statement warning that artificial intelligence is “rapidly transforming cyber‑risk.”

The alliance urged businesses to urgently adopt AI‑driven security measures before threats outpace defenses.

“Frontier AI models are expected to surpass current industry expectations, fundamentally reshaping both offensive and defensive cyber capabilities,” the statement read, echoing Warner’s sentiment that “the timeline is not years, it is months.”

Anthropic’s conflict with the administration began earlier this year over a $200 million Department of Defense contract for AI use in classified systems. The dispute became highly public when the two sides disagreed on how AI should be employed in warfare.

In February, Defense Secretary Pete Hegseth labeled Anthropic a “supply‑chain risk,” the first time the designation was applied to an American company. Anthropic has since sued the government over the label.

In April, Anthropic unveiled Mythos, a model so adept at finding security software vulnerabilities that the company warned it could pose an existential risk to digital technology. Some independent security experts echoed the concern, while others remained skeptical. Anthropic said the model would be available only to a limited set of organizations.

The NSA was among the first to receive access, placing it in the unusual position of testing a product the Pentagon had deemed a national‑security risk. Testing continued until the export‑control directive was issued this month.

Recent reports indicate that administration officials are seeking a way to resolve the dispute, including possible adjustments to the export‑control restrictions.

The White House and intelligence officials have advocated for a classified contract between Anthropic and the NSA that would allow the agency to use the company’s technology for intelligence analysis and vulnerability detection. That contract remains pending, and some Pentagon officials prefer the NSA explore alternative AI models.

Source link