AI guardrails are slowing security research

Offensive security researchers told TechCrunch that safety guardrails on OpenAI and Anthropic’s models are increasingly blocking legitimate vulnerability-discovery work, not just malicious requests, making it harder to use frontier models for the kind of exploit research that keeps software secure. The tension sits alongside both labs’ own public safety commitments, and researchers say the refusals are inconsistent enough that workarounds are becoming part of the job.