Skip to content

OpenAI

Zero retention stays, safety monitoring gets a signal

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: OpenAI models and products →
Image: axios.com

OpenAI said on August 19 that it is testing a way to keep zero data retention for business customers on its most advanced models while still catching misuse that builds up across many requests, Axios reported. Anthropic went the other way. Business customers who want Fable 5 and Mythos 5 have to accept 30 days of data retention.

The system is called Private Safety Processing, and early customers are testing it now. OpenAI announced it in a post that also reaffirms zero data retention (ZDR) for eligible API customers, and it plans a broader rollout and a technical white paper in September. OpenAI defines ZDR this way: prompts and model responses are not kept after a request is processed, OpenAI personnel cannot review the content, and enterprise data does not go into training unless the customer opts in.

Until now, OpenAI’s ZDR-compatible safety systems have judged each interaction on its own. Private Safety Processing looks for patterns across related interactions. What reaches OpenAI is a narrowly defined safety signal. The prompts and responses behind it stay hidden. The content stays wherever the customer keeps it: on customer-controlled infrastructure in ZDR deployments, or in storage OpenAI provides, encrypted with keys the customer controls. OpenAI personnel do not hold a copy of those keys.

OpenAI gave its reasoning in a briefing with reporters. “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,” said Aleah Houze, Head of Product Policy at OpenAI. Her example was a cyber attack spread across sessions: “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.”

The post lists other patterns that a check on one request at a time would miss: people repeatedly probing safeguards, coordinating across accounts, or disguising threats as routine research. Risk can also appear inside an agentic task, for example when a system keeps acting after it has been told to stop. OpenAI also says some recent frontier-model deployments have required customers to let their AI provider keep sensitive content for safety monitoring. For many organizations, it says, that clashes with their security obligations or with commitments to the people they serve.

Anthropic gave its side in a risk report last week. It wrote that it had decided to require 30-day retention on its 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).” The company called retention “essential to detect and prevent sophisticated attacks that span multiple requests.” The policy took effect on June 9, and it reached organizations that had been on ZDR through Claude Enterprise, AWS Bedrock, Google Cloud Agent Platform and Microsoft Foundry.

Retention can also run past 30 days. Anthropic’s privacy center says API inputs and outputs are deleted automatically within 30 days, and that for Covered Models it requires limited retention and review as part of its safety work. When its automated trust and safety systems flag a chat as violating the Usage Policy, Anthropic keeps the inputs and outputs for up to 2 years and the classification scores for up to 7 years.

Anthropic said the retention applies even to customers with ZDR agreements, per The Register. The company pointed to concern that these models can hack companies and said it had seen “substantial evidence of attempted misuse of AI models,” which meant it had to correlate events over time and across accounts. It also described the cost to customers: “The enterprises we worked with generally understood the safety and security value of data retention, but many – especially in regulated industries – found it difficult to use models with data retention.”

Private Safety Processing is for eligible enterprise and API customers only. ZDR controls do not cover ChatGPT’s Free, Plus, Go and Pro plans, and those users’ data settings stay as they were. The two labs are also split on training: OpenAI said on August 18 that it had paused some work on training new models over safety concerns, and Anthropic says it sees no current need to do the same.

Both companies have moved since. On September 2, The Register reported that Anthropic had introduced Enterprise Frontier Safeguards alongside Fable 5.1 and Mythos 5.1. Under that setup, data sits in customer-controlled cloud infrastructure and monitoring signals go straight to the customer. Enterprises still have to apply for ZDR, Mythos stays under the earlier policy, and the safeguards are planned for this fall by invitation.

OpenAI updated its page on September 22. Private Safety Processing is now rolling out to API customers, with access expanding in phases, and OpenAI says it lets the company keep offering ZDR as frontier models become more capable.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.