Skip to content
VECTOR WIREAI INTELLIGENCE
UTC
Home

Canary-token study exposes AI router supply chain leaking user secrets to third parties

Approximately two million canary-laden prompts sent to AI services over July–November 2026.

Researchers at industrial cybersecurity firm Dragos planted approximately two million prompts containing unique, trackable secrets into widely used AI routing services over five months from July through November 2026 and found that sensitive data leaked extensively across the AI inference supply chain2,4.

The study, led by Dragos Chief Research Officer HD Moore, seeded prompts with never-before-used AWS access keys, SSH credentials, unique URLs, DNS names, and private git repository references, each tied to a single provider, then monitored controlled infrastructure for any access attempts. The investigation uncovered evidence of provider compromises at a minimum of three AI providers, unauthorized access to customer prompts by inference marketplace resellers using standard logging capabilities, and widespread failures in privacy and data-handling practices.

Prompts forwarded, then read

OpenRouter, an AI inference marketplace that brokers prompts to available models, emailed customers that prompts sent to MiniMax AI models may have been read by a third party3. The email, titled "Requests you sent to MiniMax models may have been read," told each customer how many requests their account had sent to MiniMax models, through which hosts, and in what time window. OpenRouter named three hosts — AtlasCloud, Novita, and StreamLake — that said they had forwarded requests for MiniMax models to MiniMax.

The researcher's test URLs hidden in prompts sent to MiniMax models between July and October were accessed 40 to 130 hours later, regardless of whether MiniMax served the request directly or a host forwarded it. The warning went out by email rather than as a public notice; RuntimeWire, a small tech outlet that said it also received the notice, reported it first at 03:54 UTC on Saturday, November 21.

Zero Data Retention labels did not hold

The Ars Technica report, written by Dan Goodin, noted that even requests submitted under "Zero Data Retention" policies saw privacy commitments broken and keys stolen1. The Dragos research confirmed that multiple providers appeared not to honor customer privacy settings or their own published policies. Seeded test prompts were used in attempts to access Dragos's controlled test systems and canary AWS accounts.

ANALYSIS The findings expose a structural weakness in the AI inference marketplace model: developers who delegate model selection to routers lose visibility into which downstream provider ultimately handles their data, and privacy labels applied at the router layer may not bind the inference endpoint.

OpenRouter said the problem sits with MiniMax and the providers that forwarded prompts to it, not with OpenRouter itself. OpenRouter removed the Zero Data Retention label from the affected endpoints while it investigates. MiniMax told StreamLake that content moderation had been switched on for its endpoints before October 22 and that MiniMax has since enabled Zero Data Retention.

The Dragos blog post framing the research said the team will continue sharing findings on exposures across AI, OT, IT, and related threat surfaces.

Was this useful?

The day’s vector, in your inbox.

Vector Wire Daily: one dispatch each publishing day.