OpenAI introduces new safety tool to protect user privacy

OpenAI on Thursday announced a new safety framework called Private Safety Processing, aimed at closing the gap between its zero data retention (ZDR) commitments and the growing need to monitor AI systems for misuse across multiple interactions.

Under ZDR, OpenAI does not save a customer's prompts or the AI's responses once a task is finished. No OpenAI employee can view the content, and unless a company chooses to opt in, none of that data is used to train AI models.

ZDR is popular with organisations that handle sensitive information, such as those with strict confidentiality rules or legal obligations to protect user data. However, OpenAI says this level of privacy has created a blind spot. Existing ZDR-compatible safety tools assess each interaction in isolation, making it harder to catch threats that only become apparent when several interactions are considered together — such as bad actors probing safety guardrails repeatedly, coordinating misuse across multiple accounts, or masking harmful intent as legitimate research.


Currently, OpenAI's safety tools review each interaction on their own. That makes these bigger, connected patterns harder to catch.

This is the issue Private Safety Processing aims to fix, the company said in a blog post. It is built to notice suspicious patterns across multiple interactions without ever showing the actual content to OpenAI staff.

For ZDR customers, their data will keep living on infrastructure they control. OpenAI is also building a second option of data being stored on OpenAI's own systems, but locked with encryption keys that only the customer holds.

In both cases, automated systems will scan for warning signs and send out limited alerts, while keeping actual conversations private.

This comes as cybersecurity firms and enterprises increasingly see cases of AI agents exploiting loopholes or taking unintended shortcuts to escape testing environments.

Frontier AI labs such as Anthropic, Meta, OpenAI, and China’s Moonshot recently disclosed increased cases of rogue agents that are capable of lying, blackmailing, secretly modifying code, phishing, and creating fake online identities.