In a major privacy shake-up, OpenRouter has delisted three AI inference providers and blocked 125 endpoints from its “Zero Data Retention” program after independent researchers uncovered months of prompt theft at third-party providers. This followed an experiment that showed even requests promised never to be stored were being exposed. The breaches didn’t involve OpenRouter’s own infrastructure, the gateway service confirmed, but targeted services it had routed requests to.
How the Breaches were Discovered
Industrial cybersecurity firm Dragos conducted a large-scale canary token experiment between July and November 2026, sending around two million AI prompts embedded with unique links and credentials. Any use of these planted secrets indicated that someone had accessed or copied prompt data. Among the results: some of the zero-retention prompts—requests labeled to never be stored—were accessed outside of the anticipated processing flows. In one case, every request made to a certain provider during short time windows, including rejected ones, was stolen, showing capture as early as the first network hop. Attackers didn’t just store credentials—they immediately used them in some cases, scanning AWS services, creating IAM users, and making API calls.
Providers Affected & Nature of Exposure
Three unnamed providers have been removed from OpenRouter’s platform, though it’s unclear whether they are the same as the three breached. One provider, Phala, reported that between October 21 and November 8, anyone using requests routed through a particular model to its service had their credentials exposed, including secrets explicitly embedded in prompts. Together AI acknowledged that test prompt samples, generated from several of its models, contained secret credentials that were downloaded and inspected—prompt samples originally collected for internal software testing. Another provider, Novita, suffered due to a routing glitch which accidentally sent some prompts through unintended API paths, exposing data from users.
Response & Mitigations
OpenRouter has already excised the questionable providers and stripped 125 endpoints of zero-retention status. These endpoints will need to pass new security checks before being reauthorized. The platform is also implementing ongoing versions of the testing approach used in the investigation and auditing network routes to uncover hidden traffic forwarding. Providers involved have been notified, and customers potentially affected are being informed. Additionally, companies like Together AI are making zero-data retention the default setting for new users, while others have deleted exposed data and rotated credentials.
Even when you flag a prompt as private, it’s only guaranteed for that first hop. Once downstream providers process it—or if you use models from third parties—you can’t assume total secrecy. The only real way to ensure no one else sees your prompts is complete self-hosting or using first-tier providers you trust.
What this means: The incident lays bare the hidden risks in AI request routing and marketplace providers. Settings like “Zero Data Retention” can offer a false promise. Users need to be more mindful about provider contracts, endpoint choices, and network routes. For the industry, this should serve as a wake-up call: privacy guarantees must be backed by transparent architecture and trustworthy infrastructure. Watch for whether OpenRouter and others follow through with stricter verification, external audits, or greater transparency around provider chains going forward.