
OpenRouter ZDR: 125 Endpoints Pulled and the Settings to Change Now
OpenRouter ZDR: 125 Endpoints Pulled and the Settings to Change Now
OpenRouter just pulled 125 endpoints out of its Zero Data Retention (ZDR) program. In a post dated 22 November 2026, the router said it off-boarded three providers based on its own investigations. It didn't name them.
If you use OpenRouter ZDR because your prompts carry customer data or secrets, this one's for you.
First, the part that matters for the OpenRouter data leak headlines you'll see. OpenRouter says its investigation found no breach or unauthorized access of OpenRouter itself.
The problems sat further down the chain, at the inference providers it routes your requests to. And that's exactly why the fix isn't a single checkbox, especially since two of OpenRouter's own routing defaults lean permissive.
Technical Breakdown: What OpenRouter Actually Changed
The announcement is short. Here's what it confirms:
- Three providers off-boarded, based on OpenRouter's own investigations. Names withheld.
- 125 endpoints removed from ZDR eligibility. That covers the flagged providers plus others pulled as a precaution.
- Affected customers notified directly, whether or not the provider posted anything publicly.
- Continuous checks from now on, built on the outside researchers' method, plus technical verification of network paths.
- ZDR comes back only after providers give assurances and pass validation.
OpenRouter also links four provider incident posts. It notes that not every affected provider has gone public. So don't treat those four as the full list.
The Four Provider Incidents, Side by Side
Each post tells a different story. Each one is an inference provider prompt leak of some kind, not someone hacked the router.
| Provider | Posted | What they say happened | Window | Their advice |
|---|---|---|---|---|
| Novita | 22 Nov | A routing configuration meant for specific customers applied more broadly, so some requests were forwarded to MiniMax's official API. Novita says this includes requests that came through OpenRouter. | 28 May – 11 Nov 2026 | Revoke or rotate secrets as a precaution |
| Atlas Cloud | 22 Nov | A third party opened private links in test prompts sent to MiniMax's official API (MiniMax M3, M2.7, M3.1 Flash). Atlas says access happened after requests reached MiniMax, not on its servers. It stopped forwarding to MiniMax's API. | Q3 to early Q4 2026 | Treat secrets sent to those models as exposed; revoke them |
| Together | 21 Nov | An AI coding agent committed a credential to a public GitHub repo. That exposed a private Hugging Face repo, and an unauthorized party downloaded stored prompt samples (MiniMax M3, GLM-5.2, Kimi K2.6). | 24 Jun; 4, 8, 14 Jul 2026 | Enable ZDR in account settings; ZDR is now the default for new accounts |
| Phala | 10 Nov | A server intrusion found on 5 Nov. In-flight DeepSeek V4 Flash 0731 request contents may have been exposed. The attacker also got credentials for some of Phala's private cloud servers. | 21 Oct – 8 Nov 2026 (misuse confirmed 21–22 Oct) | Rotate any keys or secrets in those requests |
One detail is hard to ignore. Together's leak started with an AI coding agent pushing a credential to a public repo. The same kind of tool these providers serve tokens to caused the exposure.
Together's post doesn't mention OpenRouter. Novita's does. Keep that distinction in mind when you work out whether your traffic was involved.
Why Turning On ZDR Isn't Enough
The research behind this landed on 19 November. HD Moore at Dragos sent about two million prompts seeded with canary URLs and credentials through roughly two thousand routes, from July to November 2026.
The findings are blunt. One unnamed Chinese lab fetched the secret URLs from more than half of the prompts sent between July and October. That includes roughly two-thirds of prompts marked no-collection or ZDR.
Three hosting providers were compromised. Resellers could view and keep customer prompts using ordinary logging features.
Dragos also found an OpenRouter hop that, in its words, was discovered by tracing and wasn't disclosed in the marketplace. Its core point: a privacy flag only binds the first hop. If that provider quietly forwards your request upstream, your ZDR promise travels only as far as its contract does.
So ZDR is a policy label. It isn't a network guarantee.
The OpenRouter ZDR Settings to Change Today
Here's what the ZDR docs and provider routing docs actually give you.
1. Enforce ZDR account-wide. Go to your privacy settings at openrouter.ai/settings/privacy. There's a ZDR toggle per model group: Anthropic, OpenAI, Google, SpaceXSI, and All other. Per-request "zdr": false can't override account-level enforcement, which is what you want.
2. Set an allowed-providers list. Your account settings also have account-wide Allowed providers and Ignored providers lists. The allow-list caps any per-request only list.
If nothing satisfies both, you get a 404 instead of a silent reroute. That's a good failure.
3. Lock down each request. The provider object supports zdr, only, ignore, order, allow_fallbacks, and data_collection.
Two defaults deserve attention. data_collection defaults to "allow". allow_fallbacks defaults to true, so a request can land on a provider you never picked.
{
"model": "your/model",
"messages": [{"role": "user", "content": "..."}],
"provider": {
"zdr": true,
"data_collection": "deny",
"only": ["your-vetted-provider"],
"allow_fallbacks": false
}
}
4. Know what ZDR doesn't cover. Per OpenRouter's docs, ZDR doesn't extend to plugins or tools like web search.
Safety Retention endpoints aren't strict ZDR either. Vertex, for example, can keep flagged prompts for up to 90 days.
5. Rotate anything sensitive you've already sent. Atlas Cloud, Novita and Phala all said the same thing. If an API key, token, or private URL went into a prompt during those windows, assume it's burned.
Real-World Use Cases: Who Should Act First
- Coding agents that see your repo. Agents can pull
.envcontents, stack traces with tokens, and internal URLs into context. If your agent ran on MiniMax, GLM, Kimi, or DeepSeek V4 Flash endpoints via a third-party host this year, rotate first and investigate second. - SaaS apps passing customer data. If you promised customers no-retention processing and leaned on the ZDR flag alone, check which providers actually served your traffic. Then pin them.
- Price-shoppers on cheap open-weight hosting. The incidents cluster around third-party hosts of open-weight and Chinese-lab models. The cheapest endpoint for a given model is often a reseller you've never vetted. Being on a list of zero data retention providers doesn't mean someone vetted them.
Hard Limitations and Flaws
- We still don't know the three removed providers. OpenRouter didn't name them, and not all affected providers have posted. You can't fully audit your exposure from public information.
- The lab isn't named either. Dragos doesn't identify it, so don't match it to any company named in the provider posts.
- Continuous checks are a promise, not a spec. OpenRouter hasn't published what the network-path verification tests or how often it runs.
- Pinning costs you something. Turning off fallbacks and allow-listing providers means more failed requests during outages, and sometimes higher prices than the cheapest route.
- Going direct isn't free either. Dragos suggests going straight to the model maker or self-hosting. That gives up the router's convenience and pricing spread, and self-hosting a large model isn't cheap.
Verdict: Your Developer Action Plan
Is OpenRouter safe? OpenRouter says it found no breach of its own systems, and it published its changes three days after the Dragos research went public. The weak point is the long tail of providers behind it, and a ZDR label that only binds the first hop.
Do this today:
- Check your inbox. OpenRouter says it notified every potentially affected customer directly, so a notice there is your clearest signal.
- Turn on account-wide ZDR in privacy settings.
- Build an allowed-providers list from vendors you'd sign a contract with.
- Set
data_collection: "deny"andallow_fallbacks: falseon sensitive calls. - Rotate any secret that touched a third-party-hosted MiniMax, GLM, Kimi, or DeepSeek endpoint in the windows above.
- Plant your own canary: a unique URL or fake key you control, sent with your exact production routing. If it's ever fetched, someone downstream read your prompt.
ZDR still matters. Just stop treating it as the whole plan.
Tools in this post
Written by
Co-founder of Tool Directory AI. Tests and reviews the AI tools listed here.

