How OpenAI and Anthropic’s safety measures are hampering offensive security work—and why researchers say it’s counterproductive.