Skip to content
TrustList
News

OpenRouter drops three AI inference providers after research showed leaked prompts

Editorial

By TrustList Editorial

OpenRouter has off-boarded three inference providers and removed 125 endpoints from Zero Data Retention after Dragos seeded about two million prompts with canary secrets and found them used, scraped and shared.

About OpenRouter drops three AI inference providers after research showed leaked prompts

OpenRouter drops three AI inference providers after research showed leaked prompts

22 November 2026: OpenRouter, which routes requests from one API to models hosted by many inference providers, says it has off-boarded three providers and removed 125 endpoints from its Zero Data Retention (ZDR) list. The move follows research by the industrial security company Dragos, which sent about two million prompts carrying unique, trackable secrets to AI services and watched where those secrets turned up. OpenRouter says the findings showed no breach of OpenRouter itself, and that it has notified every customer who may have been affected.

Not yet independently verified. OpenRouter's actions come from its own post of 22 November 2026, and the research findings from the Dragos post by HD Moore of 19 November 2026 (both read on 22 November 2026). Neither names the three off-boarded providers or the Chinese AI lab Dragos describes. We have not read the providers' own incident posts that OpenRouter links to. We will update this when it can be confirmed, and remove this note.

Canary secrets in two million prompts

From July to November 2026, Dragos put unique URLs, DNS names and real but empty AWS keys into its prompts, so any later use of them showed that someone had kept and read the prompt. Its chief research officer, HD Moore, reports three findings:

  • Provider compromises. Data from at least three AI providers was exposed: seeded prompts were used in attempts to reach Dragos test systems and canary AWS accounts.
  • Resellers reading prompts. Sellers on an inference marketplace could log and keep customer prompts with standard logging, with no exploit needed. In one case, someone on a home broadband line opened two secret URLs about a day and a half after the prompts went through one seller's offer.
  • Privacy settings ignored. A Chinese AI lab's systems fetched the secret URLs from more than half of the prompts sent to its models, including about two thirds of those marked no data collection or ZDR, also when the requests went to US hosts that resell its models. One secret link was posted in a company chat 24 days after it was sent to a US provider with the ZDR flag. Two large US providers kept returning to secret URLs days or months after a single request.

OpenRouter's response and the provider notices

OpenRouter's post lists its steps: the three providers are gone; 125 endpoints lost ZDR eligibility, both those the researchers flagged and others removed as a precaution; and affected users were told whether or not their provider has posted publicly. It is adapting the Dragos method to test providers continuously, checking network paths for requests forwarded to third parties, and will ask for assurances before any endpoint gets ZDR status back. It links notices from Novita, Phala and Together, the last describing prompt leakage for requests in June and July 2026, and says not every affected provider has posted.

Steps for teams sending business data to AI APIs

  • Find out which models and providers your applications reach through routers, and pin the providers you trust rather than taking the cheapest route.
  • Treat ZDR and no-training flags as a promise to check, not a guarantee: ask providers for their retention terms and any onward hosting in writing.
  • Keep credentials, keys and private repository links out of prompts, and rotate any that have been sent.
  • Consider seeding your own canary values in test traffic to see whether prompts are kept.
  • If OpenRouter notified you, review what data those requests held and whether you need to report it under your data protection obligations.

Sources

Categories & features

TrustList Weekly

The week in software and IT, in one email

The news that matters to buyers, new rankings and our own research. Every Thursday, free, and easy to leave.

We will email you to confirm. Unsubscribe with one click in any issue. Privacy policy

You may be interested in