OpenRouter Drops 3 AI Providers After Prompt Leaks
OpenRouter, a service that routes AI requests across many providers, announced on November 22, 2026 that it had delisted three of them after research by security firm Dragos uncovered leaked prompts. It matters to websites, chatbots and automations that reach AI models through intermediaries.
What we know
- The research: Dragos published a study by HD Moore, its chief research officer, on November 19, 2026. Between July and November it sent about two million prompts containing decoy credentials and unique links, known as canaries, to see whether anyone used them outside the service they were sent to.
- What it found: according to Dragos, at least three inference providers (the companies that actually run the models) were breached, and its test prompts were processed alongside real customer traffic.
- The test keys: Dragos reports that a couple hundred test AWS keys (real, but tied to empty, monitored accounts) sent in late October led to tens of thousands of AWS API calls. It also says prompts sent to US-based providers reached Chinese organizations.
- OpenRouter's response (November 22): it delisted three providers without naming them and removed 125 endpoints from its Zero Data Retention (ZDR) list, some because of the findings and some as a precaution. It says it notified potentially affected customers and that its own systems were not breached.
- Phala's notice (November 10): the provider acknowledged an incident on its infrastructure. It asks users to treat as exposed any DeepSeek V4 Flash 0731 requests that OpenRouter routed to it between October 21 at 00:00 UTC and November 8 at 05:35 UTC, and to rotate any keys, passwords or other secrets they contained. The window is precautionary: confirmed misuse involves canaries sent on October 21 and 22.
- Together AI's notice (November 21): CEO Vipul Ved Prakash said a third party downloaded samples of prompts from its service from June and July 2026 (MiniMax M3, GLM-5.2 and Kimi K2.6 models) after an AI coding agent committed to a public GitHub repository a credential that gave access to the private repository where they were stored. Five of the credentials that party tested were Dragos canaries.
- What we don't know: OpenRouter does not say which three providers it removed, or whether they are the same three Dragos describes. None of the companies gives a count of affected customers.
What changes and what doesn't
What changes is how much weight a "don't retain my data" setting can carry. It is a provider policy, and Dragos warns it does not help once an attacker is already inside the provider's servers. A router may also send your request to a provider you never picked unless you pin one.
Using AI over an API is still useful. What changes is how much it pays to know who handles each request. Dragos's recommendations, also summarized by The Hacker News, include contracting directly with a top-tier provider or self-hosting the model where feasible.
How to tell if this affects you
If your site runs a chatbot, an AI plugin for WordPress or Drupal, a content generator or an automation that calls AI models over an API, check the following:
- See where your requests go: in the plugin or automation settings, find which service and API key it uses. If it is OpenRouter or another router that spreads traffic across providers, your prompts may pass through companies you did not choose.
- Check Phala's window: if you used DeepSeek V4 Flash 0731 through OpenRouter between October 21 at 00:00 UTC and November 8 at 05:35 UTC, rotate any key, password or token those prompts carried. Also check the email on your OpenRouter account: the company says it notified potentially affected customers.
- Keep secrets out of prompts: your chatbot should never receive API keys, passwords or customers' card details. If a form or workflow puts that data into the text sent to the AI, change it.
- Choose your providers: if you stay with a router, restrict in its settings which providers may serve your requests and confirm that the privacy option actually changes the route, as Dragos recommends. Avoid reseller marketplaces.
- Set your own canary: include a fake key in a test request (Thinkst's Canarytokens are free). If anyone uses it, your route leaks. That is the method Dragos used.
Related: Meta Muse: Malicious Webpage Exposed Chats in a Test
Sources
- Dragos, HD Moore, "Three breaches, endless leaks: Your AI supply chain needs canaries", November 19, 2026
- OpenRouter, "Data Processing Incidents on Inference Providers", November 22, 2026
- Phala CERT Team, "Phala November Preliminary Incident Report", November 10, 2026
- Together AI, Vipul Ved Prakash, "Security event: Prompt leakage for requests in June and July of 2026", November 21, 2026
- The Hacker News, November 22, 2026
Updates: if OpenRouter names the delisted providers or any company publishes counts of affected customers, it will be added here with a link.