OpenAI is testing a new way to monitor AI misuse without holding on to enterprise customer data. The company is previewing Private Safety Processing, a service for select customers that is designed to identify risky activity across multiple sessions while retaining none of the customer’s data.
The announcement lands in a sensitive moment for AI companies. As models become more capable, providers face pressure to prevent abuse, but enterprise customers also want strong privacy protections when they use AI systems with sensitive information.
What Private Safety Processing Is Meant To Do
Private Safety Processing is an automated safety system. OpenAI says it monitors for potential abuse while retaining none of the customer’s data. The company describes it as an expansion of Zero Data Retention, a policy already used by OpenAI and most other AI companies.
Under Zero Data Retention, agents inside the OpenAI API can check a session for abuse without the company keeping customer data. That gives AI providers a way to scan for bad activity without relying on human intervention.
The new service is meant to widen that approach. Instead of looking only at one session, Private Safety Processing can assess inputs and outputs from multiple conversations. OpenAI describes this as long-horizon safety monitoring.
That distinction matters because some misuse may not appear in a single exchange. A bad actor could divide requests across several conversations in an attempt to avoid detection. A spokesperson told TechCrunch that the system can analyze multiple conversations for signs of abuse without human review of a user’s conversations.
How OpenAI Says The System Responds
OpenAI says the monitoring is conducted by an agent. If the system is triggered, it catches interactions and analyzes them across sessions for signs of possible misuse.
When the system detects activity that needs attention, it may send OpenAI a "narrowly defined signal" about a specific type of activity. OpenAI says it can then decide whether "enforcement is necessary."
If OpenAI decides action may be needed, the company says it will contact the customer for more context or work with them on the issue. The customer may then choose whether to share data with OpenAI, according to the spokesperson.
The approach is aimed at a narrow balance: spotting potentially malicious use while avoiding routine retention or human inspection of customer conversations. For enterprise customers, that balance is central because safety reviews can become a privacy concern when large amounts of sensitive data are involved.
Why Anthropic Is Part Of The Story
OpenAI’s announcement also draws a sharp comparison with Anthropic. Anthropic recently announced a data retention policy that allows the company to keep user data for 30 days when it involves "covered models."
Those covered models include all Mythos-class models and "future models with similar capabilities," according to the source article. Anthropic’s policy was announced in July and was framed around safety, giving the lab the ability to examine potential impropriety.
Some enterprise customers have been aggravated by that approach. The concern is straightforward: companies that handle sensitive information may not want an AI lab to harbor or inspect that data, even when the stated purpose is safety.
Anthropic also largely follows Zero Data Retention, but the source notes an exception for "covered models," including Fable. For those models, its retention policy changes the privacy equation for customers using the systems.
Human Review And Customer Trust
The contrast between the two approaches is not only about retention. It is also about who may see the data and under what conditions.
Anthropic says human review of customer data can occur, but only "through a controlled access path" involving "a small set of approved reviewers." The company also says every review session is "recorded in a tamper-proof log that reviewers cannot suppress or modify."
OpenAI is positioning Private Safety Processing differently. The system is built around automated monitoring and a limited signal back to OpenAI, with customer data shared only if the customer chooses to provide it after OpenAI reaches out.
For enterprises, this difference may be important. Both companies are trying to address misuse, but the mechanics of monitoring can affect whether customers believe their conversations, workflows, and sensitive material are protected.
A Competitive Privacy Push
The timing also reflects the tense competition between OpenAI and Anthropic. Both companies are looking for advantages as enterprise AI adoption grows and customer trust becomes a business issue as much as a technical one.
The source article notes that a recent report showed OpenAI’s Q2 grew more slowly than Anthropic. Anthropic’s annualized revenue run rate is now reportedly $65 billion. Anthropic investors have said it could IPO at $2 trillion, while OpenAI is also working on its IPO.
That competitive backdrop gives privacy and safety policy a larger role. Private Safety Processing is not just a technical preview; it is also a statement about how OpenAI wants enterprise customers to think about AI safety monitoring.
The core claim is simple: OpenAI wants to detect misuse that stretches across conversations without retaining customer data or using human review of user conversations. Whether that proves persuasive to enterprises will depend on how the preview works for select customers and how clearly the company can explain the system’s boundaries.