AI models with safety guardrails, like those from OpenAI, were less direct in identifying the specific cybersecurity flaw