U.S. export controls on Anthropic’s AI models Mythos and Fable, introduced in June, have limited access for researchers seeking to test vulnerabilities, sparking concerns among offensive cybersecurity professionals about the impact of AI guardrails on their work.

In June the U.S. government imposed export restrictions on Anthropic’s Mythos and Fable models after a report said the models could be used to build malicious cyber tools. The restrictions were partially lifted; Fable 5 returned to general access on July 1 and Mythos 5 was reinstated only for vetted U.S. organizations pending review.

Both Anthropic and OpenAI provide vetted programs that give researchers reduced guardrails. Critics say the limits impede security researchers who aim to discover unknown vulnerabilities before attackers exploit them. Mark Dowd, a security researcher, said on a podcast that he feels uncomfortable with large companies deciding what is safe in security. He has sold zero‑day vulnerabilities to Western governments for years.

Chris Anley of NCC Group said prompting an AI model to exploit a bug is essential for confirming a vulnerability, but guardrails that block the prompt hinder defensive work. He likened the tool to a hammer that can be used for building or for harm. When faced with such blocks, some teams turn to open‑source models without guardrails.

Paolo Stagno, CTO of CrowdFense, said AI companies treat users like children needing supervision. He said he uses frontier models only for reverse engineering and avoids feeding exploit work into cloud models to prevent data leakage, preferring local open‑source models.

Giuseppe Cali, a zero‑day researcher, said guardrails do not impede his work because he uses AI only for code understanding, not for weaponization.

An anonymous researcher at a smartphone component manufacturer said his employer is not in Anthropic’s program, so the tools are largely unusable for security testing because the guardrails stop the model when security activity is detected.

Chris Thompson, CEO of RemoteThreat and founder of Offensive AI Con, said guardrails can be inconsistent across days even within vetted programs. He said researchers spend time negotiating with the model instead of analyzing vulnerabilities, and that many are moving to Chinese open‑source models such as GLM that have no usage restrictions.

Thompson called for AI labs to open their programs, provide responsible access, and hold abusers accountable, arguing that defenders risk falling behind as attacks scale rapidly.