How AI guardrails are impeding the work of offensive cybersecurity researchers

Cybersecurity researchers report that safety guardrails in AI models from OpenAI and Anthropic are increasingly blocking their efforts to identify software vulnerabilities and build proof-of-concept exploits. While these measures are intended to prevent malicious use, experts argue that over-sensitive filters hinder legitimate security testing by refusing to generate code or analysis required for defensive research. This creates a friction point where security professionals find it harder to use AI tools for discovering flaws before bad actors do.