OpenAI previews zero-retention safety system as Anthropic requires data logs
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Aïda Amer/Axios
OpenAI said Wednesday that it believes a new technique will allow it to safely serve its most advanced models to businesses without needing to retain their data.
Why it matters: Rival Anthropic has instituted a 30-day retention policy for business customers who want to use Fable 5 and Mythos 5.
Driving the news: OpenAI is testing "Private Safety Processing" with early customers. The system is designed to identify misuse patterns across related interactions while preserving zero data retention, or ZDR, protections.
- OpenAI says its systems would send the company a narrowly defined safety signal, without exposing the underlying prompts or responses.
- Customer data can stay on customer-controlled infrastructure, or be stored by OpenAI with encryption keys controlled by the customer.
- The company plans a broader rollout and technical white paper in September.
What they're saying: OpenAI says it needs some amount of context over time to detect risks, but that its new system lets it do so without requiring access to sensitive customer data.
- "We're seeing with more capable frontier models that often risks are emerging not just by looking at one single prompt and response pair, but when you look over time at multiple interactions," Aleah Houze, Head of Product Policy at OpenAI said in a briefing with reporters.
- "So, an example of this would be somebody might be asking about the weakness in a company's software in one conversation, and then later in another conversation they might ask about remote access or what security tools can detect."
- Seeing these signals in a broader context might help the company detect a cyber attack.
The other side: Anthropic says retaining data is necessary for security.
- "We have recently announced our plan to require 30-day data retention on our most capable models—a decision we believe will be unpopular with customers who have come to expect zero retention, and pose real risks to our business success (especially if competitors do not follow), but which we believe is essential to detect and prevent sophisticated attacks that span multiple requests," the company wrote in a risk report last week.
State of play: It's another area where the two rivals are diverging as models enter a new era of risk and capabilities.
- OpenAI said Tuesday that it has paused some work in the training of new models to deal with safety concerns, while Anthropic says it doesn't see a current need to do the same.
Zoom in: The new safety system is currently designed for eligible enterprise and API customers, not people using OpenAI's paid or subscription consumer ChatGPT plans.
- OpenAI's ZDR controls don't apply to Free, Plus, Go and Pro users and their existing consumer data settings remain unchanged.
