AI Guardrails Hamper Offensive Cybersecurity Research

Frontier AI models can accelerate reverse engineering, vulnerability validation and exploit development, but the same capabilities can also be abused in cyberattacks. Anthropic and OpenAI have therefore imposed safeguards on high-risk requests while offering vetted researchers access through Anthropic’s Cyber Verification Program and OpenAI’s Trusted Access for Cyber. Security practitioners say the dual-use nature of their work makes clean distinctions difficult: instructions that help confirm and fix a flaw can also reveal how to exploit it.

TechCrunch reported on July 24, 2026, that researchers at NCC Group, CrowdFense and RemoteThreat encounter refusals and shifting restrictions that waste time and push some teams toward locally run or Chinese open-source models such as GLM. The U.S. imposed export controls on Mythos and Fable in June; Fable 5 returned to general access on July 1, while Mythos 5 was restored only for vetted U.S. organizations during a government review. No financial terms were disclosed.