OpenRouter removed three AI inference providers from its platform on November 22, 2026, after security firm Dragos found attackers had been quietly intercepting customer prompts at some of its upstream providers since mid-July. OpenRouter also pulled 125 endpoints out of Zero Data Retention (ZDR) eligibility and says it has notified every customer whose requests may have been exposed. The company is clear that its own infrastructure was not breached. The exposure sat with specific providers it routes traffic to.
How Dragos Caught the AI Providers Leaking Data
Dragos ran a straightforward trick at scale. Between July and November, researchers seeded roughly two million prompts sent through various AI gateways with fake, trackable secrets: unique links, planted AWS keys, that kind of thing. Then they watched for anyone using them.
The worst case they found involved a couple hundred planted AWS keys sent to a single provider across a few windows in late October. Every one of them got used within hours. Some of the stolen requests had already been rejected by the provider before a model ever ran on them, which tells you the copying was happening right at the front door, not somewhere deeper in the pipeline. In two of the three confirmed breaches, requests flagged as Zero Data Retention got stolen anyway.
One attacker pointed an AI-driven penetration testing tool at the stolen keys just to confirm they still worked, then used them for tens of thousands of AWS API calls: listing S3 buckets, hunting secrets stores, trying to add IAM users. Dragos also flagged a separate issue on an AI inference marketplace, where resellers could read and keep customer prompts just by leaving standard request logging turned on. No exploit needed.
What OpenRouter Changed
Per its own announcement, OpenRouter off-boarded three providers based on its investigation into Dragos’s findings and pulled 125 endpoints out of ZDR eligibility, some because they were directly flagged, others as a precaution. It notified every potentially affected customer, including ones whose providers haven’t posted anything publicly.
Going forward, OpenRouter says it’s turning Dragos’s canary methodology into an ongoing, continuous check rather than a one-time audit, and auditing its own network paths to catch cases where a provider quietly forwards traffic to a third party without saying so.
What This Means if You Build on AI APIs
A “zero retention” toggle only governs the first hop. If a provider reroutes your traffic, or gets compromised itself, that setting does nothing to stop the copy. If a client product of yours calls models through a router or an aggregator, it’s worth knowing which actual provider is sitting behind that brand name on any given request, not just trusting the label on the dashboard.
It’s also a reminder that whatever an AI agent can read, it can send. Credentials dropped into prompt context travel with the request, through every hop that touches it. Planting a fake credential as a tripwire, the same trick Dragos used, is cheap enough that any team shipping agents in production can do it themselves with something like Thinkst Canarytokens.
This is the kind of review we build into AI development work for clients now by default: pinning providers instead of accepting a default routing pool, auditing exactly where prompt data can end up, and keeping real credentials out of anything an agent can read. It fits a pattern we’ve been tracking for months of attackers going after the AI supply chain itself rather than the application code sitting on top of it, see our coverage of the backdoored Microsoft Agent Framework package and the API design choices that make agent integrations safer to use.

