OpenAI zero-data monitoring counters Anthropic 30-day retention
This digest was compiled by AI from multiple sources — links to the originals are below.
OpenAI’s Private Safety Processing system counters Anthropic’s 30-day retention policy by monitoring AI misuse without storing customer data. The automated system scans multiple conversations and sends only narrow risk signals to OpenAI servers. Companies Glean, Databricks, Abridge and Microsoft are testing the technology.
Key Facts
- Private Safety Processing is an automated system that monitors AI misuse across multiple sessions while retaining none of the customer’s data, OpenAI says.
- Anthropic’s July data-retention policy keeps user sessions for covered models for 30 days, including all Mythos-class models and future models with similar capabilities.
- The system scans interactions through local or encrypted channels and sends only narrow signals about risk type and degree to OpenAI servers, without exposing message content.
- Glean, Databricks, Abridge and Microsoft participated in testing the Private Safety Processing tool.
- A full launch of the technology and publication of a technical paper are scheduled for September.
Private Safety Processing
OpenAI offers select enterprise customers an automated misuse-monitoring system called Private Safety Processing. The system upholds Zero Data Retention, meaning no customer query or API response is stored on OpenAI servers. It scans inputs and outputs across multiple conversations, not just single sessions, to detect malicious use spread over time. When triggered, an agent analyzes interactions across sessions and sends the server only narrow signals about the type and degree of risk. This monitoring operates without human review of a user’s conversations and without giving OpenAI staff direct access to message content.
Anthropic’s Retention Policy
Anthropic announced in July that it may keep user data for 30 days for covered models, including all Mythos-class models and future models with similar capabilities. The policy was designed to allow safety analysis of potential misuse, but it has aggravated some enterprise customers handling sensitive data. Anthropic largely follows Zero Data Retention for non-covered models, with retention applying only to covered models. OpenAI’s Private Safety Processing extends ZDR to long-horizon monitoring across multiple sessions, in contrast to Anthropic’s retention of covered-model conversations.
Customer Testing and Launch
Companies Glean, Databricks, Abridge and Microsoft participated in testing the Private Safety Processing tool. OpenAI says the system detects malicious AI use that unfolds over multiple sessions, such as requests spread out to avoid detection. A full launch of the technology and publication of a technical paper are scheduled for September.
2 sources
OpenAI zero-data monitoring counters Anthropic 30-day retention
