OpenAI published a new security post on February 25, 2026 and linked it to the latest installment in its ongoing threat-reporting work on "malicious uses of AI." The company says the report is built around case studies from investigations into misuse of its models, and that the purpose is not just to announce enforcement actions but to explain how abuse appears in practice when investigators follow an operation end to end.

The core point is that threat actors usually do not run AI-enabled operations inside a single model or a single product. OpenAI says malicious activity is typically combined with older tools such as websites, social media accounts, and other operational infrastructure. It also says abuse is not always limited to one AI provider, citing a Chinese influence operator as an example of how actors can move across models and platforms while pursuing one campaign objective.

What OpenAI is emphasizing

AI misuse often sits inside a broader workflow rather than replacing the rest of the workflow.
Cross-platform investigation matters because attackers may distribute tasks across multiple services.
Publishing case studies is meant to help other companies and defenders spot similar patterns earlier.

That framing matters for enterprises and governments that are putting frontier models into production. The post suggests that safety work cannot stop at prompt filtering or model moderation alone. Teams also need account monitoring, link and infrastructure analysis, incident response playbooks, and information sharing across platforms if they want to detect coordinated abuse before it scales.

OpenAI says it has been publishing these threat reports for two years and presents the new document as part of a broader effort to help the industry and wider society identify and avoid threats. Even in summary form, the message is clear: the operational context around a model interaction is now a central security surface in its own right.

#abuse-detection

OpenAI publishes a threat report on malicious AI campaigns spanning models, websites, and social platforms