A group of cybersecurity executives, researchers and policy experts is asking the U.S. government to lift export controls placed on Anthropic’s Fable and Mythos large language models, arguing that the restrictions could weaken defenders more than adversaries.
In an open letter addressed to Commerce Secretary Howard Lutnick and National Cyber Director Sean Cairncross, the signatories said the policy should be reversed and replaced with a more open process for evaluating AI-related security risks. The letter frames the issue as both a national security concern and a question of how AI safety decisions should be made.
The group says advanced AI systems are already changing cybersecurity by making it easier to find software flaws and generate exploits. At the same time, the signatories argue that Anthropic’s Mythos-class models are capable of assisting with vulnerability discovery and exploit development, but are not the only tools with those abilities. Many security professionals, the letter says, already rely on other commercial and open-source models for audits and red-teaming.
A central argument in the letter is that Anthropic has built strong safeguards into the Fable model to curb offensive cyber use. Those protections, the authors say, were conspicuous enough to draw jokes in the security community when the model launched. Despite that, they contend that export restrictions now limit access for legitimate users such as developers and security teams.
The letter says AI tools are important for identifying weaknesses in both newly written software and older systems that still power much of the internet and enterprise infrastructure. Restricting access to those tools, the signatories argue, could slow efforts to fix vulnerabilities before attackers exploit them.
The authors also point to competition from abroad, saying Chinese open-weight models are only months behind leading U.S. systems, based on public benchmarking. They argue that unpublished capabilities may exist as well, raising the risk that American defenders could be left behind if access to top-tier models is curtailed without a strong justification.
The letter disputes the premise that the Anthropic models offer uniquely dangerous capabilities. According to the signatories, the underlying research that prompted the export action focused on whether a person could use the model to determine if code was insecure. They say that ability is useful for secure coding and should not automatically be treated as an offensive cyber function.
The coalition also says similar results can be achieved with other leading models from U.S. and Chinese developers. On that basis, it argues that the government’s action is unusually broad and has introduced market uncertainty while limiting access for legitimate defensive work.
Beyond the specific case of Anthropic, the letter calls for any future AI cyber regulation to follow a transparent, scientific and democratic process. The signatories say assessments should involve industry and academic input, be enforced fairly, and allow time for remediation. They also say any restrictions should be limited to what is strictly necessary to protect the public.
The signers include a wide range of people from the cybersecurity field, among them company founders, chief information security officers, researchers and former government officials. The letter states that affiliations are listed for reference only and do not represent organizational endorsement.
The appeal adds to a broader debate in Washington over how to manage advanced AI systems that can both strengthen cyber defenses and increase offensive capabilities. For now, the letter makes clear that many in the security community believe the balance should favor wider access for defenders, not tighter restrictions.