OpenRouter says it has removed three inference providers from its platform and stripped 125 endpoints of Zero Data Retention eligibility after security researchers found evidence that some providers’ handling of customer prompts did not match their stated protections. The company also notified users it considered potentially affected.

The response follows canary-token testing across the AI inference supply chain. Such tests place unique, trackable markers inside prompts and monitor whether those markers are later accessed. A signal outside the expected processing path can show that prompt content was retained, copied or exposed somewhere between a customer and the hardware serving a model.

OpenRouter acts as an intermediary that gives customers access to models operated by multiple hosting companies. That arrangement can simplify model selection and billing, but it also means a request may pass through several organizations. Privacy labels therefore depend not only on the router’s own systems but also on whether every selected provider enforces the promised policy. Zero Data Retention, commonly shortened to ZDR, is intended to limit storage of prompts and responses after processing.

In its incident notice, OpenRouter said the findings concerned specific inference providers and did not demonstrate a breach or unauthorized access of OpenRouter itself. That distinction matters: the company is reporting action against members of its provider network, not a compromise of its routing platform. It said it had been working with researchers and providers to understand the results and encourage remediation.

The company’s measures go beyond removing the three providers. Reclassifying 125 endpoints means customers can no longer select those routes under OpenRouter’s ZDR eligibility until their status changes. OpenRouter said it contacted affected users even when a provider had not issued its own public incident report, and it is urging providers to publish accounts of what happened. The company plans to maintain a list of provider security announcements as more become available.

The episode highlights the difficulty of enforcing data promises in a rapidly changing marketplace. A privacy control shown at the customer-facing layer is only as strong as the downstream systems that receive the request. OpenRouter said it plans to incorporate similar testing techniques into its oversight of the network and require providers seeking access to meet a higher standard. For customers handling proprietary code, internal documents or personal information, the immediate lesson is to assess the full inference route rather than treating a single platform label as proof of end-to-end protection.