Key takeaways

  • OpenRouter says it does not store prompts or responses unless you opt in, but every provider it routes to sees them in full.
  • A Dragos study published November 19, 2026 found prompts stolen from three hosting providers and privacy flags ignored, including on ZDR requests.
  • Enforcing ZDR only, pinning providers with fallbacks off, and putting a spend limit and expiry on every key cut most of the risk.

OpenRouter can be used safely for most work if you change its defaults, but it is not a private pipe, and a ZDR label is not a guarantee. OpenRouter itself does not store your prompts unless you opt in, according to its data collection docs. The risk sits with the dozens of companies it forwards your prompts to, and research published on November 19, 2026 shows that some of them leaked prompts, ignored privacy settings and lost credentials to attackers.

What matters is which providers your requests reach, and you control that. The steps below come from OpenRouter's docs as of November 21, 2026.

What OpenRouter stores and what it doesn't

OpenRouter is a router: it sends each request to a provider that hosts the model, based on price, speed and reliability. That provider sees your full prompt and the response.

OpenRouter says it does not store prompts or responses unless you opt in to logging them for debugging, or to letting it use them for a 1% discount. Both are off by default.

One exception: on October 27, 2026, OpenRouter announced it would screen all requests for child sexual abuse material, including traffic under zero data retention, and may keep flagged content for review and legal reporting. It does store metadata for every request: token counts, latency, the model used and timestamps. Its privacy policy, last updated August 31, 2026, says it does not use API customers' inputs or outputs to train models.

The real risk is the providers

Each provider has its own data policy. Some keep prompts for a while, some may train on them, and some keep nothing. OpenRouter labels every endpoint with one of three tiers in its zero data retention docs:

Tier What it means Trained on?
ZDR No routine storage after processing No
Safety Retention Prompts may be kept, only to detect abuse No
Retention and Training Stored, and may be used for training under the provider's policy May be

Those labels come from providers' terms and agreements. OpenRouter's terms of service say it "makes no representation or warranty regarding any model provider's data handling."

Two defaults matter. The request setting data_collection defaults to allow, which permits providers that store data and may train on it. And with ZDR on, self-serve accounts still allow Safety Retention endpoints unless you choose otherwise.

What the November 2026 canary research found

HD Moore, chief research officer at Dragos, published a study on November 19, 2026. From July to November his team sent about two million prompts to AI services, each carrying unique fake secrets such as AWS keys and URLs. Any later use of a secret showed where the prompt had gone.

What turned up, according to the report:

  • Attackers stole data from three hosting providers. In the largest case, a couple hundred fake AWS keys sent to one provider in late October were stolen and used for tens of thousands of AWS API calls.
  • Some resellers in model marketplaces listed popular models cheaply, logged the prompts and passed the secrets on.
  • A Chinese AI lab's systems fetched secret URLs from more than half of the prompts sent to its models, including about two-thirds of requests marked as no data collection or ZDR.

It mentions OpenRouter once, as a hop found by tracing traffic behind a marketplace. Moore's point is that a privacy flag "only binds the first hop" and does nothing against an attacker already inside a provider. More on the findings is in the news story on the research.

What OpenRouter told customers about MiniMax

On November 21, several OpenRouter customers posted on Hacker News an email they said they received from OpenRouter, titled "Requests you sent to MiniMax models may have been read." According to the copies posted, OpenRouter said:

  • A researcher's test URLs in prompts to MiniMax models were opened by a third party 40 to 130 hours later, so prompts and outputs may have been read.
  • AtlasCloud, Novita and StreamLake told OpenRouter they had forwarded those requests to MiniMax.
  • OpenRouter blamed MiniMax and the forwarding providers, removed the ZDR label from those endpoints while it investigates, and was still waiting for MiniMax's own account.

OpenRouter had posted nothing public about it by November 21. Its public ZDR endpoint list that day still had MiniMax models from hosts such as DeepInfra and Together, but none from AtlasCloud, Novita, StreamLake or MiniMax itself.

How to make OpenRouter safer

These settings are in OpenRouter's privacy settings, its provider routing docs and its key settings.

Turn on ZDR and choose "ZDR only"

In privacy settings you can enforce ZDR for five model groups: Anthropic, OpenAI, Google, SpaceXSI and all other models. Then pick ZDR only instead of Allow Safety Retention to exclude endpoints that may keep prompts. Most Google models drop out, because Google has required Safety Retention on Vertex since October 18, 2026.

You can also enforce it per request. The request setting can only add ZDR, never switch off an account-wide rule:

{
  "model": "your-model",
  "messages": [],
  "provider": {
    "zdr": true,
    "data_collection": "deny"
  }
}

ZDR covers model inference only, not plugins or tools such as web search.

Select your providers

To pick providers yourself, list them in order and set allow_fallbacks to false, so OpenRouter fails instead of quietly trying someone else. To allow a fixed set, use only. Both settings also exist account-wide, and the account list acts as the ceiling.

"provider": {
  "order": ["provider-a", "provider-b"],
  "allow_fallbacks": false
}

Copy exact provider slugs from each model page. Moore's advice is the same: choose upstream providers yourself and avoid reseller marketplaces.

Block a provider

To block a provider on one request, add it to ignore. To block it everywhere, add it to the ignored providers list in privacy settings. The two lists merge.

Limit what a stolen key can spend

OpenRouter recommends a credit limit on every key, because a key with no limit "lets a leaked key or a runaway agent spend your entire balance, including any auto top-ups." Limits can reset daily, weekly or monthly. The Security settings page, added on September 28, 2026, flags keys with no limit, keys that never expire and idle keys. You can also set a maximum key lifetime, after which OpenRouter rejects any key that never expires, old ones included. IP allowlists need an Enterprise plan.

OpenRouter emails you if it finds an exposed key, but its terms make you responsible for all charges on your credentials unless OpenRouter's own breach caused them.

Use guardrails for teams

Guardrails apply budgets, provider allowlists, ZDR and a Sensitive Info filter that redacts personal data to specific keys or members. A guardrail does nothing until you assign it.

Finally, keep real secrets out of prompts. No router setting helps once a compromised provider has read them.

Is OpenRouter legit

Yes. It is a real company that routes requests to 400+ models from more than 80 providers, and its customers include NVIDIA, Zoom and Lovable, according to Stripe's announcement of August 19, 2026 that it agreed to buy OpenRouter. OpenRouter said then that it expected the deal to close "in the coming weeks" and that its name, product and roadmap would stay the same. Neither company's announcement page mentions a closing date as of November 21.

Is OpenRouter free

Creating an account and an API key costs nothing, but most models are paid per token. OpenRouter passes through provider prices and charges a 5.5% fee, $0.80 minimum, on credit purchases, per its FAQ. Free variants ending in :free allow 50 requests a day, or 1,000 after you buy 10 credits, per the limits docs.

Update, November 22: OpenRouter said it off-boarded three providers and removed 125 endpoints from zero data retention eligibility after the Dragos findings, and that the tests showed no breach of OpenRouter itself. It linked incident notices from Novita, Phala and Together AI and says it will now run similar checks continuously. Our news report lists the affected models and dates.

Frequently asked questions

Is OpenRouter legit?

Yes. It routes requests to 400+ models from more than 80 providers, and Stripe announced on August 19, 2026 that it agreed to acquire it.

Is OpenRouter free?

Signing up and creating an API key are free, but most models are paid per token, with a 5.5% fee when you buy credits. Free model variants ending in :free are capped at 50 requests a day, or 1,000 a day after you buy at least 10 credits.

How do I block a provider on OpenRouter?

Add its slug to the ignore list in the provider object of a request, or to the ignored providers list in your privacy settings to block it on every request.

How do I select a provider on OpenRouter?

List providers in the order field and set allow_fallbacks to false, or use the only field to allow a fixed set. Exact provider slugs are on each model page.

Sources

  1. Data CollectionOpenRouter
  2. Zero Data RetentionOpenRouter
  3. Provider RoutingOpenRouter
  4. Provider LoggingOpenRouter
  5. Zero data retention endpoints listOpenRouter
  6. AuthenticationOpenRouter
  7. Security settingsOpenRouter
  8. GuardrailsOpenRouter
  9. LimitsOpenRouter
  10. Frequently Asked QuestionsOpenRouter
  11. Privacy PolicyOpenRouter
  12. Terms of ServiceOpenRouter
  13. CSAM Screening Across OpenRouterOpenRouter
  14. OpenRouter is joining StripeOpenRouter
  15. Stripe agrees to acquire OpenRouterStripe
  16. Three breaches, endless leaks: Your AI supply chain needs canariesDragos
  17. Tell HN: OpenRouter prompts to Minimax may have been readHacker News
  18. Data Processing Incidents on Inference ProvidersOpenRouter