OpenAI Previews Private Safety Processing for Zero Data Retention
OpenAI has previewed Private Safety Processing, a system designed to identify misuse across several interactions simultaneously, ensuring zero data retention. The company emphasizes that its staff never access the content of these interactions.
Key Points:
- Zero Data Retention (ZDR): OpenAI promises not to retain customer data for its next-generation models, and this includes prompts and model responses.
- Private Safety Processing: This new system aims to detect harmful intent that may only be apparent when considering multiple related interactions together.
- Four Promises: OpenAI guarantees that:
- They do not keep prompts or model responses once a request is processed.
- Their personnel cannot access this content for review.
- Enterprise data does not train the models, unless the customer opts in.
- Models are not trained on enterprise data.
- Competitor Contrast: Anthropic, however, recently announced a policy requiring 30 days of log retention for its most advanced models, arguing that's necessary to catch attacks spanning multiple requests.
Why Private Safety Processing Matters:
OpenAI highlights that existing ZDR systems evaluate each interaction in isolation, missing potential threats that emerge over time and across related interactions.
The post lists several types of misuse patterns it aims to catch:
- Repeatedly testing safeguards.
- Account coordination for malicious purposes.
- Disguising harmful intent as legitimate research.
- Agent drift (when an AI system deviates from its intended purpose and continues acting even after a stop command).