跳到正文
@rohanpaul_ai· @rohanpaul_ai · X·· 2026-08-20AI 评分45
AI 导读

OpenAI 预览 Private Safety Processing,让符合条件的零数据保留(ZDR)客户在请求处理后仍可使用前沿模型,且 OpenAI 不保留其提示词与回复。

正文

OpenAI’s latest move against Anthropic: promising not to retain customer data.

They are making Zero Data Retention a frontier-model feature through Private Safety Processing.
So eligible Zero Data Retention customers can keep using frontier models without OpenAI retaining their prompts or responses after the request is processed.

"Private Safety Processing" will detect risky patterns across interactions without giving OpenAI employees access to the underlying customer content.

To understand the context, each individual prompt can look harmless, while the sequence of prompts reveals a dangerous plan. For example, asking separately about a pathogen, how to increase its spread, and how to evade detection may each look legitimate alone, but together they can show harmful intent. Existing ZDR safeguards mostly see each request in isolation, so they can miss that larger pattern.

Hence, Private Safety Processing instead analyzes related activity while content stays on customer-controlled infrastructure or in OpenAI storage encrypted with customer-controlled keys.

When it detects a risk, OpenAI says it receives a narrow category-and-severity signal, while personnel still cannot access the underlying prompts or responses.

So OpenAI is moving a hard safety problem from reading single prompts to deciding which interactions belong together.

Private Safety Processing is supposed to catch patterns across related interactions while staff remain unable to see the underlying content. To do that, the system still needs some durable way to link events through identifiers or customer-provided grouping. Attackers can evade weak linkage by fragmenting activity across accounts. If providers broaden the linkage, customers inherit more metadata that can reveal relationships across otherwise separate interactions.

Anthropic, by contrast, requires 30-day retention for Claude Fable 5 and Claude Mythos 5, which it designates Covered Models. Anthropic still offers ZDR for eligible API use elsewhere, so its 30-day rule does not cover every business request.

引用@OpenAI@OpenAI
We will continue to offer Zero Data Retention for frontier models. As AI takes on longer, more autonomous work and delivers greater value to businesses, safety systems also need to identify risks across related interactions. To help address those risks, we're previewing Private Safety Processing, which is designed to improve safety without giving OpenAI personnel access to the underlying content.
在 X 查看被引用的帖子

来源:@rohanpaul_ai · x.com