Allegations that some security measures in Anthropic’s Claude Fable 5 model can be bypassed have accelerated discussions between the company and the US government. During meetings in Washington, the possibility of reopening the model for consumer use was discussed.
The US government maintains the view that some protection layers of the model can be overcome, potentially granting access to more advanced cybersecurity capabilities. Anthropic argues that these concerns are exaggerated.
Meetings Held at the Department of Commerce
Negotiations took place in working group meetings organized at the Department of Commerce. Participants included government researchers and officials from the National Cyber Director’s Office; Commerce Secretary Howard Lutnick also joined via conference from the G7 Summit.
Anthropic was represented by co-founder and chief computing officer Tom Brown and head of external relations Sarah Heck. Logan Graham, head of the company’s advanced security testing team, and senior security researcher Nicholas Carlini also traveled to Washington to participate in the talks.
A company spokesperson stated that both sides are working to resolve the issue quickly. The White House declined to comment on the matter.
Security Concerns Led to Restriction Decision
The discussions focused on the differences between Claude Fable 5 and the company’s Mythos model. Fable 5 is presented as a version limited by certain protections in cybersecurity, biology, and chemistry; if these protections are bypassed, it is believed that access to the more powerful capabilities of the Mythos model could be gained.
Last week, the US government took action for the first time due to security concerns. It was reported that Amazon CEO Andy Jassy directly called Treasury Secretary Scott Bessent, which heightened concerns within the administration. Subsequently, White House officials tasked the NSA with conducting a security vulnerability assessment.
The NSA reportedly concluded that the protections could indeed be bypassed, leading the administration to decide to restrict the model. During this process, Lutnick also held talks with Anthropic CEO Dario Amodei.
After Anthropic suspended model access for all users over the weekend, Lutnick had multiple phone conversations with Brown and Heck. The Department of Commerce expressed openness to finding a way to reopen the model to users once security concerns are fully addressed.
The timing of the crisis coincided with Anthropic’s ongoing dispute with the Pentagon over the military use of some AI models. As a result, the discussion evolved beyond a purely technical security issue to a broader topic affecting the company’s relations with the government.
Open Letter and Industry Reactions
Some cybersecurity researchers argued that the government’s move was unnecessary and opposed the decision with an open letter. The signatories noted that while the company’s models are strong in finding vulnerabilities and generating exploits, they are not uniquely so in this field.
Researcher Katie Moussouris stated that the current findings do not fully prove a jailbreak. According to Moussouris, security protections are not absolute barriers for a skilled attacker but rather obstacles that slow down progress.
Investors reportedly spent the weekend assessing how this tension might impact the company’s future. Some investors believe the US government is specifically targeting Anthropic, while others think a similar response might not occur if another company released a comparable model.
This development serves as a warning to other labs producing advanced AI models. It is expected that companies will engage with the White House earlier and more transparently before launching models.
Cohere CEO Aidan Gomez said the events demonstrate that the US government could take such steps across the industry. Gomez emphasized that companies should not be caught unprepared for this reality.
