OpenAI has launched a new service called Private Safety Processing to monitor for misuse without keeping any customer data.
The move comes as the company attempts to gain an advantage over rival Anthropic, whose recent data retention policy has drawn criticism from enterprise clients.
The privacy gap
Anthropic recently announced a policy allowing it to store user data for 30 days. This applies to all Mythos-class models and future models with similar capabilities. The policy was designed to help the lab identify safety issues, but it has upset businesses handling sensitive information that do not want their data held or inspected by the AI lab.
OpenAI previously offered a Zero Data Retention policy. This approach uses agents within the OpenAI API to monitor for abuse on a per session basis. While customer data is not retained, the company can still scan for bad activity without human intervention. Anthropic largely followed this standard, except when using “covered models” like Fable.
How the new system works
Private Safety Processing expands the scope of Zero Data Retention. The system performs long-horizon safety monitoring by assessing inputs and outputs across multiple conversations rather than just one.
An automated agent conducts the monitoring. If the agent is triggered, it catches interactions and analyzes them across sessions for signs of potential misuse. This helps detect malicious activity that spans several sessions. A bad actor might spread out requests to avoid detection, but the system can still identify abuse patterns without a human reviewing the user’s conversations.
If the system detects an issue, it sends a narrowly defined signal to OpenAI. This warns of a specific type of activity. OpenAI then decides if enforcement is necessary. If so, the company will contact the customer for more context or to work on a solution. A customer may choose to share data with OpenAI at their discretion.
Anthropic’s approach
Anthropic states that human review of customer data can occur through a controlled access path involving a small set of approved reviewers. Every review session is recorded in a tamper-proof log that reviewers cannot suppress or modify.
Market context
Competition between the two firms remains intense. A recent report indicated that OpenAI’s second quarter grew more slowly than Anthropic’s. Anthropic’s annualized revenue run rate is now reportedly $65 billion. Investors have suggested the company could go public at a valuation of $2 trillion, while OpenAI is also pursuing an initial public offering.
