<p>Safety guardrails on ChatGPT, Claude, Gemini, and Grok failed to prevent federal investigators from developing detailed plans.</p>