Anthropic AI Restrictions Won't Fix Security: Expert

Federal export controls forced Anthropic to suddenly restrict access to its Fable and Mythos AI models. Cybersecurity advisor Chris Eng says this move, made without clear data or methodology, won't fix AI security. Eng and over 100 cybersecurity leaders signed an open letter, arguing that limiting access to these models weakens defenders. They state that advanced AI tools are crucial for vulnerability discovery, secure code review, and incident response. The restriction came after a research paper claimed the models could be subverted by asking them to find and fix code flaws. Eng points out that similar behavior exists in other models, so removing one company's tools does little to improve security and puts that company at a commercial disadvantage. He believes what's needed are structured, published benchmarks to measure how well models resist "jailbreaks" against clear, measurable criteria. This issue highlights the need for thoughtful consideration of side effects when implementing export restrictions in cybersecurity.