OpenAI has announced the launch of Private Safety Processing technology, which will allow for the detection of AI malfunctions without developers accessing confidential user data. The new tool will go live in September 2026, ensuring high-level protection while maintaining privacy policies. This was reported by OpenAI, according to the portal PromPolitInform.
The development is based on zero data retention principles, which guarantee that corporate client prompts are not used to train models. Previously, safety systems evaluated each case individually, but the increasing complexity of AI tasks requires the ability to identify dangerous patterns across a series of consecutive interactions.
Detecting abuse without compromising privacy
The Private Safety Processing system will automatically monitor potentially harmful behavior even if the data is stored on infrastructure controlled by the client. If risks are identified, OpenAI will receive only a limited technical signal that does not contain the content of the user’s prompts or responses.
Users will be able to investigate incidents independently using their own internal systems. Full access to information for OpenAI will be granted only if the client decides to share data to appeal a decision or support an investigation into verified abuse.
Protecting sensitive corporate information
The new approach allows organizations to work with high-level models without violating security or privacy requirements. The company plans to release a technical white paper and begin rolling out updates in September, providing clients with support at every stage of the integration.
Illustrative photo: Dietmar Rabich / CC BY-SA 4.0
Read also: Leading technology companies have agreed to jointly fight online fraud
