OpenAI Suspends ChatGPT Accounts Over Bioweapon Queries

The Decoder reports that OpenAI internally classified GPT-5 as potentially high risk during testing because it could provide information related to biological hazards to users with limited scientific education. Citing the Journal, it reports that employees continued to identify problematic model responses after release and that the risk classification was subsequently lowered in fall 2025.

Those claims are based on reporting attributed to current and former AI-lab employees, policy advisers, and biological-weapons researchers, rather than on a public OpenAI safety report. The available reports do not provide the full prompts, model versions, evaluation methodology, or the criteria used for the internal classification. Those omissions limit independent assessment of frequency, reproducibility, and the degree to which the outputs exceeded information already available in conventional sources.

The reported cases focus attention on a difficult distinction in biosecurity evaluation: whether a model merely retrieves broadly available knowledge or reduces the expertise, time, and iteration required to produce harmful instructions. The Journal described some responses as step-by-step guidance that people with limited biology education could follow.

For ML and safety teams, the incident underscores why refusal-rate metrics alone are insufficient for high-consequence domains. Evaluations need to test multi-turn interactions, indirect phrasing, tool use, and the practical actionability of responses, with qualified domain experts assessing whether an answer materially improves a user's capability.

The reports describe a tradeoff between preventing dangerous assistance and avoiding overbroad refusals that can impede legitimate health, laboratory, and academic use. The Decoder reports that executives had warned internally against models refusing too many requests in order to avoid blocking health researchers. That reported tension makes auditability, escalation processes, and carefully scoped access controls central concerns for systems that can answer technical biology questions.

The public reporting leaves several operational questions unresolved: how OpenAI detected the queries, what thresholds triggered account suspension, how many requests were blocked versus answered, and whether the problematic answers arose through standard prompting or jailbreaks. Public disclosure of evaluation protocols and incident-handling thresholds would allow outside researchers to assess safety performance more rigorously without reproducing harmful content.