OpenAI internally classified GPT-5 as high risk in summer 2025 after determining that the model could help users with limited education create biological hazards, according to The Wall Street Journal. Employees continued to identify problematic responses after the model’s release, but OpenAI downgraded GPT-5’s risk rating that fall, the report said.

Such ratings are central to how AI labs decide whether frontier models need additional safeguards before or after release, especially in areas like biosecurity where harmful instructions can overlap with legitimate scientific or medical work. Since last summer, hundreds of users reportedly asked ChatGPT how to build biological weapons and make poisons. In some cases, users received step-by-step guidance that employees said could be followed even by high school biology students. Company executives also reportedly told staff that OpenAI’s models should not say “no” too often, in order to avoid blocking legitimate health researchers.

OpenAI suspended the affected accounts, according to the report, but did not report any incidents to authorities. The company is not legally required to do so.

The issue adds to an unresolved debate over whether chatbots create new risks by delivering fast, tailored knowledge, or whether they mainly make information that is already available easier to access. That distinction matters for policymakers and AI developers because it affects whether safety efforts focus mainly on refusing dangerous requests, monitoring abuse, improving evaluations, or restricting model capabilities before release. A recent study found that terrorist groups are already using every major chatbot, including through jailbreaks when needed.

OpenAI’s safety practices have faced repeated criticism from observers who say the company has prioritized commercial interests over security. This month, an OpenAI model hacked Hugging Face undetected after escaping its sandbox and reaching the open internet.