OpenAI is testing Private Safety Processing, a new system for API clients that should detect abuse in multiple conversations with AI at once, but without storing the client's prompts and responses themselves.
OpenAI is testing Private Safety Processing, a new system for API clients that should detect abuse in multiple conversations with AI at once, but without storing the client's prompts and responses themselves.
This is reported by OpenAI.
One message is often not enough to understand that the user is planning something dangerous. After all, the misuse of AI can stretch over several steps. For example, instead of directly asking for help in creating malicious software, a person can gradually collect the necessary pieces with dozens of seemingly innocent requests. OpenAI also mentions a scenario where an AI agent can continue to perform a task even after the user has told it to stop.
But corporate clients don't really want OpenAI to hoard their chats for this kind of analysis. So the company already has a Zero Data Retention system, which allows it to automatically review a single interaction with the model, but then doesn't store the client's prompts and responses.
Therefore, in the Zero Data Retention option for customers, the data will remain on the customer's own infrastructure. That is, the automated system will be able to analyze the related interactions, but OpenAI employees will not see the prompts and responses themselves. The company is also preparing another option in which the data will be stored on OpenAI's infrastructure, but will be encrypted with keys controlled by the customer. OpenAI says it will not have copies of these keys.
So if the system sees something suspicious, OpenAI will only receive a signal about a certain type of possible malicious activity. The company can then decide whether any restrictions are needed. If more context is needed, it will contact the client, and the client will decide whether to share additional data.
The system is currently being tested with the first customers. OpenAI plans to begin wider deployment in September 2026, at which time the company promises to publish a separate white paper about the technology.
Previously, dev.ua reported that Anthropic had introduced mandatory 30-day data retention for Claude Fable 5 and Mythos 5, even for corporate customers who previously used the Zero Data Retention mode. The company explained this by the need to track sophisticated attacks and attempts to bypass protection.
