Private Safety Processing lets OpenAI review related API requests for safety without retaining your prompts or responses. Records stay in your environment, where an automated safety agent reviews them without OpenAI personnel being able to access your content.

See how it works and connect your cloud storage:

Ayyy I recognised a goat here! @codestantine

shut up till you release a model sincerely ~Everyone 💖

支持你们，希望你们的产品越来越好。💪

The key word is "related". One request on its own rarely looks like misuse, a pattern across many calls does. So they review requests together, and the trade is that no OpenAI staff read them, an agent does, and the records stay in your environment Most of what I send to models is storyboards and render scripts for Claude Code films, so this matters less to me. For teams putting contracts or medical records through an API it's probably the first question, before anything about model quality

靠

划重点：ZDR+PSP 把抽检的 prompt/response 加密写进你控的存储（S3/Azure/GCS/OCI），OpenAI 只留索引元数据+存储引用。审核在硬件证明、禁人工访问的 Runtime 跑，明文只出有界安全信号，TTL 30 天。想问：一开是整个 project 全量走 PSP，还是可按模型/端点豁免？

Shouldn’t this just be a built in standard? The beauty of Astra 6 is no need for us non-software engineers to have to figure things out or use Cursor etc. The Ai model should handle all complexities. Human ideas are our “IP”, rare and valuable, should be protected from the start. @sama @gdb @thsottiaux @OpenAI

openai says a bot reads your requests so no human has to. best coworker ever, zero gossip

AI safety without sacrificing data privacy. That's an interesting direction. The future of AI needs both security and user trust.

so my agent gets a hall monitor that lives on my infra. whose tokens does it burn

if nobody at openai can see the content, appealing a false positive from that safety agent gets awkward fast

我们想要的是一个新的模型，不是这个东西

this matters a lot for anyone building on health or finance data. real q tho, do devs get told why a request got flagged or is it just vibes

and when the agent flags it, who looks?

even the moderation logs are going private…..

the prompts still leave the building to get answered though, this settles the retention question, not the one most of my clients actually ask

I never liked sending prompts out just for a safety check

haha finally — safety review without keeping the data is the bit everyone skips 😅

Come on dude stop with these nothingburger posts and release a fucking model soon

I thought it was.....

Records staying in your own environment is a big deal tbh. That's the bit that's stopped a lot of companies going all in on the API

my local agents finally have an OpenAI safety agent to argue with.

double encryption keeps content locked to customer keys

just put the model in the api bro

新模型什么时候来？

The diagram ends at "limited safety signals" and doesn't say what's in them. If a signal can carry a summary of the flagged prompt, the content still leaves your cloud, just shorter.

So an AI agent checks the AI agents. Who reviews the safety agent though? What a time to be alive lol