Offensive security researchers say AI guardrails block their work

Cybersecurity researchers who hunt for vulnerabilities and build exploit tools say OpenAI's and Anthropic's safety guardrails are getting in the way of legitimate defensive work. TechCrunch interviewed several researchers about how model refusals and restrictions slow down vulnerability research.

Why it matters: This highlights a recurring tension in AI safety design: broad refusal policies meant to stop malicious use also blunt tools for the defenders who rely on the same techniques. Expect vendors to face pressure to build verified-researcher exceptions, similar to ongoing debates around red-teaming and CTF use.