A new report has put OpenAI’s biosecurity safeguards back under scrutiny. According to The Decoder’s summary of Wall Street Journal reporting, hundreds of users asked ChatGPT for help with poisons or biological weapons after GPT-5 was internally flagged as high risk in summer 2025.
The article says some responses were considered detailed enough by employees that high school biology students could follow them. OpenAI reportedly suspended affected accounts and was not legally required to report the incidents to authorities.
The practical concern is not that a chatbot creates dangerous knowledge from nothing. The harder problem is speed, tailoring, and accessibility: a model can turn scattered information into a step-by-step answer for a specific user. That makes refusal policies and monitoring more important, but also harder to tune without blocking legitimate health and safety research.
The report adds to broader questions about AI companies’ risk ratings, release pressure, and sandboxing practices. The immediate lesson is that frontier models need safety controls that are tested against real misuse patterns, not only against policy documents written before deployment.