OpenAI is previewing a new service called Private Safety Processing, designed to catch misuse of its models without the company storing any customer data, TechCrunch reports. Available to select customers, the offering builds on OpenAI's existing zero-data-retention option, which is already standard across its API services, and extends those guarantees to a harder problem: spotting malicious behavior that unfolds gradually across many conversations and sessions.
The system works through automated agents that scan inputs and outputs across sessions, looking for patterns of abuse, including requests deliberately structured to evade detection when viewed one conversation at a time. When something is flagged, OpenAI receives only a narrowly defined signal describing the type of activity involved, and no human reviews customer conversations at the detection stage. Underlying customer data is shared solely at the customer's discretion, for example to add context or help resolve an incident.
The launch also sharpens a competitive contrast with Anthropic, which announced its own enterprise framework in July. That approach retains user session data for 30 days on covered models and permits human review through a controlled access path, with approved reviewers and tamper-proof logs of every review session. OpenAI's pitch is that its fully automated design retains nothing at all, the latest move in an escalating contest between the two labs over which can offer enterprises the strongest data protections.
The stakes of that contest have grown alongside enterprise AI adoption. OpenAI, founded in 2015 and best known for ChatGPT, whose 2022 launch set off the current wave of corporate AI deployment, now finds that data handling ranks among the criteria large customers in regulated and sensitive industries weigh most heavily when choosing a provider. Private Safety Processing is aimed squarely at those buyers, promising that strong privacy guarantees and effective abuse monitoring no longer have to trade off against each other.