OpenRouter removes 125 endpoints from Zero Data Retention eligibility
OpenRouter says it off-boarded three providers and removed 125 endpoints from Zero Data Retention eligibility. The company says it notified all potentially affected customers.
Key takeaways · 4
- 01
OpenRouter says it off-boarded three providers after its investigations.
- 02
The company removed 125 endpoints from Zero Data Retention eligibility, including some as a precaution.
- 03
OpenRouter says the research found no breach or unauthorized access of OpenRouter itself.
- 04
Researchers found secret URLs accessed when prompts went to MiniMax directly or through AtlasCloud, Novita or StreamLake, OpenRouter said.
What the tests found
Researchers sent nearly two million prompts containing unique, trackable secrets to AI providers, placing each secret in exactly one prompt sent to one provider.[1] The test included functioning AWS credentials tied to empty accounts, SSH passwords, unique URLs and DNS names, and private Git repository references.[1] The experiment found evidence that attackers compromised at least three AI providers and siphoned off prompts.[1] The stolen bait keys were used tens of thousands of times in AWS API calls.[1] These findings concern the test and its planted secrets; they do not by themselves establish the amount of real customer data exposed.
Why routing and privacy labels matter
AI routers send prompts to available inference engines, sometimes choosing based on price or task suitability.[1] Although prompts are encrypted in transit, they are decrypted at each hop inside an inference reseller network.[1] The research found secret URLs fetched from more than half of prompts, and from roughly two-thirds of requests marked no data collection or Zero Data Retention.[1] The reporting also notes that a privacy flag does not protect a prompt if an attacker is already inside a provider.[1] A label therefore cannot, on its own, establish that every part of a service chain is secure.
OpenRouter’s response
OpenRouter said researchers found its planted URLs accessed when prompts went to MiniMax directly or through AtlasCloud, Novita, or StreamLake.[1] It removed a Zero Data Retention label from MiniMax-model endpoints.[1] The company says it off-boarded three providers and removed 125 endpoints from Zero Data Retention eligibility, including precautionary removals.[2] OpenRouter said the research found no breach or unauthorized access of OpenRouter itself.[2] It also said it notified potentially affected customers, including users whose providers had not posted public incident reports.[2]
Checks before endpoints return
OpenRouter says it is adapting the researchers’ testing approach to run checks on an ongoing basis and checking network paths for requests and responses that might be forwarded to third parties.[2] It says endpoints removed from Zero Data Retention eligibility will need assurances and validation before that eligibility is restored.[2] Separately, researchers found that prompts routed through US hosting providers reselling a Chinese lab’s models triggered the same URL scraper as prompts sent to the lab.[1] A Git client on another cloud later tried to clone private repositories named in prompts using those models.[1]
The findings show that prompts can pass through routers and inference resellers, where they are decrypted at each hop, and that secret URLs were fetched from requests marked no data collection or Zero Data Retention. Professionals can assess the full chain handling a prompt, including which endpoints receive sensitive data, how requests may be routed, and what evidence a provider requires before restoring a privacy designation.
Why it matters
AI training for your team
kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.
How this developed
22 November 2026
OpenRouter said it off-boarded three providers and removed 125 endpoints from Zero Data Retention eligibility, including some as a precaution, and notified potentially affected customers.
21 November 2026
Canary test found prompt theft across AI services
Sources
- Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced. - Ars Technicaarstechnica.com
- Dragos: canaries catch data theft at three AI inference providerstheinjection.dev
- Data Processing Incidents on Inference Providersopenrouter.ai
- OpenRouter Drops 3 Providers After Researchers Catch Months of Prompt Theft at AI Hosting Firmsthehackernews.com
