When you use AI APIs today, your prompts and responses often get stored by the provider—sometimes for safety review, sometimes for model training. OpenAI just changed that for enterprise users: their new Zero Data Retention (ZDR) policy means your API calls leave no trace after processing, unless you specifically allow it.
This matters if you’re building anything with sensitive data: medical apps, financial tools, or business systems handling trade secrets. Until now, using frontier AI models meant accepting that your prompts might be reviewed or used to train future models. Now you can opt out entirely.
What actually changed today
Two key shifts happened:
- No default retention: Enterprise API customers get automatic zero data retention—OpenAI won’t keep your prompts or responses after processing them, unless you opt in to sharing.
- Private Safety Processing: A new system detects harmful patterns across multiple API calls without humans ever seeing your content.
For example, if someone tries to bypass safety filters by splitting a harmful request across five API calls, the system can connect those dots automatically. It works by analyzing metadata and patterns, not by storing your actual text.
How this affects your projects
If you’re learning to build with AI, here’s what this means for your work:
- Healthcare or finance apps become more viable—you can now process patient records or financial data without that information being stored by OpenAI.
- Internal business tools gain privacy—your company’s strategy discussions or product plans won’t linger in AI provider logs.
- Agent workflows get safer—long-running AI agents that make many API calls can be monitored for drift without exposing their full conversation history.
Important
This doesn’t mean your data is magically secure. You still need to handle API keys carefully, audit your own logs, and follow standard data protection practices. Zero retention at OpenAI’s end doesn’t eliminate risks in your own code.
The storage options available
You control where your data lives during processing:
| Option | Where data lives | Encryption | Who holds keys |
|---|---|---|---|
| Customer infrastructure | Your servers/cloud | Your choice | You |
| OpenAI infrastructure | OpenAI servers | AES-256 | You |
The second option—OpenAI infrastructure with customer-held keys—is new. It’s designed for teams who want to use OpenAI’s scale without giving up control. Your data gets encrypted before leaving your network, and OpenAI can’t decrypt it.
Why patterns matter more than single prompts
Early AI safety systems looked at each API call in isolation. That fails when:
- Someone slowly probes for vulnerabilities across dozens of calls
- An agent gradually drifts from its original task
- Multiple users coordinate an attack from different accounts
Private Safety Processing creates anonymized risk signals—like “possible prompt injection attempt” without showing the actual prompts. OpenAI gets just enough information to block abuse, not enough to reconstruct your data.
What stays the same
Some things haven’t changed:
- Free-tier users still have data retained by default
- Non-enterprise API plans still use data for model improvement unless you opt out
- All users still get basic per-request safety filtering
This is primarily an enterprise feature, though the techniques may trickle down to other tiers later.
When you might opt back in
Zero retention isn’t always better. Consider allowing some data storage if:
- You need OpenAI’s help debugging odd model behavior
- You’re fine with your non-sensitive data improving future models
- You want human review of safety flags (you can share select snippets case-by-case)
The opt-in is granular—you can allow retention for specific projects while keeping others fully private.
How this compares to other providers
Most AI companies retain data by default:
| Provider | Default retention | Opt-out available | Cross-request safety |
|---|---|---|---|
| OpenAI (Enterprise) | No | N/A (already off) | Yes (Private Safety Processing) |
| Anthropic | Yes | Partial | Per-request only |
| Google Gemini | Yes | No | Per-request only |
Microsoft’s Azure OpenAI Service offers similar enterprise controls, but without the pattern-detection across calls.
What to do next
If you’re building with OpenAI’s API:
- Check your plan: Only enterprise contracts get automatic ZDR today
- Review storage options: Decide whether to use your infrastructure or OpenAI’s encrypted storage
- Monitor safety signals: Set up alerts for any risk flags in your API responses
- Audit old projects: Anything built before today may have different privacy settings
For personal projects, the free tier still works the same way—this mainly affects professional systems handling sensitive data.
OpenAI’s move shows a clear trend: as AI becomes infrastructure, it needs to disappear into the background while keeping its safety rails. That’s good news for anyone building real-world applications where privacy can’t be an afterthought.
Read next
- How Stampli cut launch time by 68% with AI agents
- How Google’s Agent Development Kit Secures AI Agents for Beginners
Want to try all of this hands-on? Start with the free Vibe Coding 101 course.
Source
Based on OpenAI’s announcement, “Offering Zero Data Retention for frontier models”. Written for people learning to build with these tools.