Is zero data retention enough to keep AI prompts private?
Some stolen requests carried a zero-retention flag, and a Chinese AI lab fetched secret links from prompts marked no data collection. OpenRouter has since cut 125 endpoints from its ZDR list.

No. Dragos sent about two million prompts carrying fake secrets to AI services between July and November 2026 and found attackers inside at least three providers. Some of the stolen requests were marked zero data retention (ZDR), and the flag did not stop the theft. On November 22, OpenRouter removed three providers and 125 endpoints from its ZDR program.
What did Dragos find when it seeded two million prompts?
Dragos chief research officer HD Moore and his team sent nearly two million prompts from July through November 2026. Each carried unique URLs and canary credentials, which are fake secrets nobody has a reason to use, so any use shows the prompt was read. Dragos published the results on November 19.
Dragos's evidence points to prompt theft at three or more AI providers. The first incident dates to mid-July, and the providers learned of it only when Dragos told them. In the largest breach, a couple hundred fake AWS keys were each sent once, inside a few windows in late October, and every request Dragos sent in those windows was stolen. Some keys were taken from requests the provider turned away before any model ran, which points to copying at the provider's front door. The attackers tested the keys, in some cases within hours, then listed S3 buckets and searched Secrets Manager. Across the breaches, the stolen keys drove well over ten thousand AWS API calls.
Why did the zero data retention flag not protect the prompts?
Dragos writes that a privacy flag governs only the first hop and does nothing once an attacker is inside the provider. In two of the three breaches, the stolen requests included some marked ZDR.
The flag also failed with no breach at all. Systems at a Chinese AI lab opened the secret URLs in over half the prompts Dragos sent to its models from July to October. That included about two-thirds of the requests marked no data collection or ZDR. U.S. hosts that resell that lab's models triggered the same scraper, so a customer who chose a U.S. provider to keep data in the U.S. did not get that. In one case a seeded link was posted in a Feishu chat in China 24 days after the request, which Dragos says carried the ZDR flag and went to a U.S.-based provider.
Resellers were a third route. Sellers on an inference marketplace read buyers' prompts through ordinary request logs, which needed no exploit. Prompts for the same model sent through a different seller were never touched, so the leak depended on which seller handled the request.
What has OpenRouter changed so far?
On November 22 OpenRouter said it dropped three inference providers from its platform and pulled 125 endpoints from ZDR eligibility, according to UrduPoint and Unite.AI. The 125 include providers the researchers flagged and others pulled as a precaution, so the number is not a count of confirmed breaches. OpenRouter said it notified customers who might have been affected, and that the testing showed no breach of OpenRouter itself. It will restore the ZDR label only after it has assurances and validation, and it is adapting the researchers' tests to run on an ongoing basis.
An earlier step came on November 21. UrduPoint, summarizing an Ars Technica update, reports that OpenRouter had already emailed customers to say MiniMax prompts might have been read. The planted URLs were opened when prompts went to MiniMax directly and through the resellers AtlasCloud, Novita and StreamLake, and OpenRouter removed the ZDR label from MiniMax endpoints.
Which three providers were breached?
Dragos has not named them, and OpenRouter's public list does not say which three it removed. That list links security notices from Novita, Phala and Together AI, and not every affected provider has published one, according to Unite.AI. Dragos says one of the late-October providers has published a preliminary security notice. The provider hit first, in July, has not disclosed its incident publicly, and Dragos says stolen data from it was still leading to login attempts as of late October. Until names are confirmed, review every route your prompts take.
What should teams that route prompts through AI gateways do today?
- List every router, gateway and reseller between your application and the model. Dragos says each hop reads the full prompt and the full response.
- Check whether any route used an endpoint OpenRouter removed from ZDR eligibility, and read its security notices.
- Choose the upstream providers yourself instead of taking the default pool, and test that your privacy settings change where requests go.
- Stay away from reseller marketplaces. Dragos says part of that capacity runs on stolen keys.
- Rotate any key or password that has ever appeared in a prompt sent through a router. Dragos notes that AI agents send whatever they can read, credentials included.
- Plant a fake credential where your agents will pick it up, and handle any use of it as a security incident. A free Thinkst Canarytoken is enough to start.
- Turn on secret filtering where a provider offers it. Dragos says it can't hurt but is not a panacea, because plenty of sensitive data does not look like a key.
- Going directly to a first-tier provider gives the smallest exposure, though it can still leak. Self-hosting is the only route where nobody else reads your prompts, and Dragos says it mainly makes sense for the most sensitive work.
Sources: Dragos, UrduPoint, Unite.AI, DEV Community. Related: Notion AI chats leaked via Fireworks, Railway breach: rotate secrets.