OpenRouter Drops Three Providers After AI Prompt Exposure: What Customers Should Check

OpenRouter removed three providers and 125 endpoints from zero-data-retention eligibility after research found prompt exposure. Here is what customers can check.
Rows of server racks inside a modern data center, used to illustrate AI infrastructure capacity
Photo by Brett Sayles via Pexels.

OpenRouter has removed three AI inference providers from its platform and stripped 125 endpoints of zero-data-retention eligibility after security researchers found evidence of exposed prompts and failures to honor data policies.

The AI routing service announced the changes on November 22, saying it had notified potentially affected customers. Its investigation found no breach or unauthorized access of OpenRouter itself. The 125 endpoints include some removed as a precaution, so the number should not be read as a count of compromised services.

The action follows a Dragos investigation that sent roughly two million prompts containing unique, traceable links and credentials to AI services from July through November. Researchers observed evidence that data at at least three providers had been exposed, as well as instances in which providers did not appear to follow data-handling settings.

How a prompt can travel beyond the model

OpenRouter connects an application to a choice of models and the companies that host them. A model name and a hosting provider are not necessarily the same thing: a request can pass through a router, an inference provider and, in some cases, another company to which that provider forwards it. Each hop can affect where a prompt is processed and who can access it.

Dragos planted different credentials or links in individual requests and watched for subsequent use. In its largest observed breach, a few hundred test AWS keys were stolen from requests sent during short windows in late October. Some came from requests the provider had rejected before running a model, suggesting the copying occurred at the provider’s front door. The researchers observed attempts to list storage buckets, search secret stores and create access for continued use.

Some stolen requests in two other breaches carried a zero-data-retention designation, according to Dragos. That setting governs a provider’s handling of data; it is not a guarantee against an intruder who can access a system while a request is being processed.

What Together AI disclosed

One provider, Together AI, disclosed a separate prompt-leakage incident on November 21. It said an unauthorized party downloaded samples stored for software testing in a private Hugging Face repository after a credential was exposed in a public GitHub repository. Together attributed the commit containing the credential to an AI coding agent.

The samples involved requests to MiniMax M3, GLM 5.2 and Kimi 2.6, primarily on June 24, July 4, July 8 and July 14. The attacker searched the downloaded samples for other service credentials and tested them, including five Dragos canaries, Together reported. The company said its review of the affected contents and customer notifications was ongoing.

Together says those samples were authorized for testing and that it has made zero data retention the default for new accounts, among other changes. The OpenRouter announcement lists Together and other provider disclosures but does not identify the three providers it removed. Customers should not assume the named disclosures correspond to that removal list.

What customers should do now

  • Check notifications and routing history. OpenRouter says it notified potentially affected customers. Review the provider, model and time of relevant requests in available records, and contact the service if the notice or history does not establish exposure.
  • Identify sensitive inputs. Consider prompts, system instructions, uploaded context, code and tool outputs that may have contained credentials, personal information or confidential business data. Avoid putting additional secrets into a support request.
  • Rotate exposed credentials. If a potentially affected request contained a live key or token, revoke and replace it, then examine the relevant provider’s access logs for unexpected use. Follow your organization’s incident and notification process where warranted.
  • Review retention and routing settings. Check whether the exact endpoint remains eligible for zero data retention and whether fallback routing or subcontracting can send requests elsewhere. Ask providers what retention covers, where processing occurs and how compliance is verified.
  • Reduce future exposure. Remove secrets from prompts where possible, use narrowly scoped, short-lived credentials and limit what connected agents can read. For particularly sensitive workloads, assess direct provider arrangements or self-hosting against the operational costs and security responsibilities involved.

OpenRouter says it is adapting the researchers’ tests for ongoing checks and examining network paths for forwarding. Together’s investigation remains open, and the public disclosures do not establish the full number of affected customers or the contents of every exposed request.

Leave a Reply

Your email address will not be published. Required fields are marked *

Previous Post
Server racks in a data center used to illustrate enterprise networking and security systems

MONAI Label 0.8.6 Fixes Exploited Flaw: What Server Operators Should Do

Related Posts
Liquid cooling hoses and server infrastructure inside an NVIDIA AI factory reference design

NVIDIA Rubin Pushes AI Data Centers Toward Hotter, Drier Cooling

NVIDIA says its Rubin-generation AI infrastructure can run fully liquid-cooled servers with 45°C coolant, cutting facility cooling water use from conventional tower-based levels to near zero in favorable climates. The design is a real shift for AI factories, but it does not erase the water tied to power generation, chip manufacturing, or local data center siting fights.
Read More