Nearly 150 cybersecurity leaders have signed an open letter urging the Trump administration to reverse restrictions on Anthropic's AI models, warning that the crackdown could weaken American defenses against hackers.
The dispute centers on Anthropic's Fable and Mythos models, which the administration moved to restrict after concerns emerged about potential Chinese access to the technology. But security experts say the real problem is the precedent being set: that American AI companies face punishment for building tools that help defenders identify and patch vulnerabilities.
The threat, they argue, is that frontier AI developers may now strip defensive capabilities from their models to avoid government retaliation. That would leave U.S. cybersecurity teams without tools their adversaries continue to develop and deploy.
"They've set a precedent that American models can't do defensive security research," said Alex Stamos, the former Facebook security chief who organized the letter. "This is closer to China than what I recognize as the United States, and personally I see this as a huge threat to American dynamism."
The escalation began when Amazon flagged concerns about Fable and Mythos, reportedly prompting alarm in the administration. Anthropic engaged a leading zero-day researcher with deep ties to the Defense Department to assess those concerns. That expert, who helped create the Pentagon's bug bounty program and sits on multiple government advisory boards, is now being portrayed by the administration as a "radical Democrat."
What Amazon reportedly found was a jailbreak allowing Fable to write what security researchers call "proofs of concept" - demonstrations of how vulnerabilities actually work. Katie Moussoulis, CEO of Luta Security, reviewed Amazon's findings and concluded the issue involved narrow prompts designed to support defensive work, not mass exploitation.
Anthropic had tested both models extensively before release with internal teams and outside researchers. The company has also noted that perfect jailbreak resistance is not currently achievable for any AI provider, so it focused on making jailbreaks either narrow in scope or expensive to produce.
The real concern among security leaders is straightforward: if companies believe they'll face government action for releasing models useful to defenders, they'll simply remove those capabilities. One security expert put it bluntly on social media: "There is no fix that wouldn't render the model less useful for cyber defenders. No new frontier models can be developed or released if this is the administration's best take."
The paradox troubles researchers. Chinese AI developers and state-backed hacking groups are unlikely to handicap their own tools in response to American pressure. That leaves U.S. defenders at a disadvantage, potentially losing access to capabilities their adversaries actively use.
The administration is also standing up a new AI vulnerability clearinghouse through a recent executive order, which would triage reports of jailbreaks and other AI model threats. Questions remain, however, about available cybersecurity talent in the White House following recent departures and the effective sidelining of the nation's top cyber agency.
Author James Rodriguez: "The administration is trying to prevent one hypothetical disaster and creating another in the process - this kind of policy whiplash is exactly how defensive capabilities get strangled."
Comments