Two Million Tripwires: What a Canary Experiment Revealed About the AI Router Economy
Your prompt to an AI model may pass through several companies that read it in plaintext, log it, and, in some cases, hand it along. Researchers at security firm Dragos just measured how bad that supply chain is by planting nearly 2 million digital tripwires in it.
HD Moore, Dragos's chief research officer, assembled unused secrets: AWS access keys tied to empty accounts, SSH passwords, unique URLs, DNS names, and private git repository references. Each secret went into exactly one prompt, sent to one provider, through the growing market of AI routers and gateways like OpenRouter and Vercel, which broker prompts to whichever inference provider is cheapest or best suited. Because Moore controlled the infrastructure behind every secret, any hit told him precisely which link in the chain leaked.
The results, published Thursday and reported by Ars Technica, describe an ecosystem with more holes than policies. "What we found revealed security breaches, prompt theft, and widespread failures to protect customer data," Moore wrote.
Three kinds of failure
Breached providers. Three separate hosting providers in the reseller network were compromised. At one, attackers reused a couple hundred of the planted AWS keys. Every request sent during a few short windows in late October was stolen, including requests the provider rejected before any model ran, which places the attacker at the provider's front door rather than deeper in the inference stack. In two other breaches, some stolen requests had been marked Zero Data Retention. The stolen keys were used tens of thousands of times in AWS API calls: the parties browsed S3 buckets, secrets managers and parameter stores, tried to enable hosted models in Bedrock, and added their own IAM users. Secrets sent to two of the three providers were used together within hours from the same rented server, pointing to one adversary.
Resellers running the theft themselves. Some providers, identified by sending the same model requests through different sellers, lured traffic with low prices and simply logged requests. When credentials resurfaced only on the path through one seller, the reseller was the culprit, not the model provider.
Flag violation at scale. The clearest scraper case involves a Chinese AI lab whose systems fetched the secret URLs from more than half of the prompts sent to its models between July and October, including roughly two-thirds of requests marked "no data collection" or ZDR. The scraper typically appeared two or three days after the request. Crucially, it also fired on prompts sent through several US hosting providers that resell the lab's models, meaning the prompts flowed back to the lab. A customer paying a US provider to keep data in the US did not get what they paid for. A git client on a different cloud later began trying to clone private repositories referenced in prompts, but only prompts that used that lab's models. One US host crawled a secret URL more than two months after a single inference call.

The market is reacting
OpenRouter on November 21 emailed customers that some of their prompts may have been read, citing the research. The exposure involved prompts sent to MiniMax, the China-based model provider, whether sent directly or through three resellers, AtlasCloud, Novita and StreamLake. OpenRouter has removed the "Zero Data Retention" label from MiniMax endpoints.
Moore's remediation advice is grim reading for anyone who picked a router for price. The safest option is self-hosting, which is expensive. The next best is buying directly from a first-tier provider with one contract and one privacy policy. Users should decline default provider pools and pick upstream providers explicitly. None of it guarantees privacy; it narrows the chain you have to trust.
The deeper point is structural. Prompts are encrypted in transit but decrypted at every hop, and each router, pool operator and host plausibly keeps a copy for as long as its policy says, if indefinitely. Canary tokens are the first method that turns that architecture into a measurable leak rate. The rate, on two million samples, was not low.
Sources
- Dan Goodin / Ars Technica: https://arstechnica.com/security/2026/11/a-leaky-ai-supply-chain-is-exposing-secret-user-data-at-an-alarming-rate/
- HD Moore / Dragos research write-up: https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens
- OpenRouter customer notification, via Ars update and Hacker News: https://news.ycombinator.com/item?id=50585962
- Techmeme pick: https://www.techmeme.com/261122/p3
