According to a report from Tech Times, OpenAI identified that its GPT-5 model posed a bioweapon-related risk and downgraded that risk rating before the model was released.
The claim is a serious one because of how AI companies police themselves. Leading labs, OpenAI included, publish internal safety frameworks that assign risk levels to a model's capabilities in sensitive areas — chemical, biological, nuclear and cyber threats among them. Those ratings are not just paperwork: they determine what safeguards get built, what a model is allowed to say, and in the strictest cases whether it ships at all. A rating is effectively the company grading its own homework, and the grade decides the consequences.
So the allegation described by Tech Times cuts at the credibility of that whole system. If a lab can register a concern and then revise the number downward on its own authority, the framework functions less as a brake than as a formality. That matters most for biological risk, the category safety researchers most often single out, because the worry is not that a chatbot invents a new pathogen but that it lowers the skill barrier — walking a determined non-expert through steps that would otherwise require specialist training.
It is worth being precise about what is not yet established here. The available source is a single report; it does not spell out who made the decision, what evidence prompted the change, what the original and revised ratings were, or how OpenAI has responded. Those details are the difference between a routine reassessment after further testing and something closer to a governance failure.
Why it matters: AI safety in the United States currently rests largely on companies rating their own products, and this story goes directly to whether that honor system can be trusted.