Anthropic has sent senior technical staff to Washington, D.C., to meet with Trump administration officials this week in hopes of ironing out issues after the White House raised national security concerns about the company’s top artificial intelligence models, a source close to Anthropic told the Washington Examiner.
The White House on Friday imposed export restrictions on Anthropic’s Fable 5 and Mythos 5 AI models after suspicions arose that a group linked to China had accessed the Mythos 5 model. In response, Anthropic took the programs completely offline.
Recommended Stories
The source close with Anthropic said company officials have met virtually with Trump administration officials every day since they were first contacted about concerns of an alleged way to “jailbreak” the models’ security features over the weekend.
However, White House AI czar David Sacks said that Anthropic CEO Dario Amodei was less than cooperative when the Trump administration asked the company to resolve the “jailbreak,” which led the administration to “reluctantly” place restrictions on public access to the models.
“In the past, Anthropic has always said that safety must be top priority and taken super seriously,” Sacks said. “In this case, Anthropic prioritized the continued offering of the consumer model over safety.”
Amodei attempted to diffuse the administration’s concerns in three separate phone calls with several officials, according to Politico, arguing that the security bypass that happened did not amount to a “jailbreak” that would permit the model to be used without the guardrails put in place by Anthropic.
White House officials were reportedly unconvinced by Amodei’s argument and cited Amazon recently finding a method to jailbreak the Mythos 5 model, as reported by the Wall Street Journal.
Anthropic released a statement on Friday announcing that all access to Fable 5 and Mythos 5 would be suspended to any foreign national following the government’s raising a “national security concern.” The statement reflected what Amodei reportedly told administration officials about the scale of the breach.
“We reviewed a demonstration of this specific technique being used to identify a small number of previously known, minor vulnerabilities,” Anthropic said. “These vulnerabilities all appear relatively simple, and we have found that other publicly-available models are able to discover them as well without requiring a bypass.”
Anthropic explained that it had only received “verbal evidence” of a “potential narrow, non-universal jailbreak,” and said that it disagrees with the reaction to block the models deployed to the public.
AMAZON RESEARCHERS USE ANTHROPIC AI TO FIND CYBERSECURITY WEAKNESSES
“If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers,” the statement said.
War Secretary Pete Hegseth placed the White House restriction against his own department’s decision to end the Pentagon’s contract with Anthropic, saying it was “the right move.”
