
Two million canary prompts expose leaks across the AI inference supply chain
Dragos researchers tracked about two million unique canary prompts from July through November 2026 and reported signs of compromise at three providers, prompt access by marketplace resellers, and apparent failures to honor privacy controls. Phala's incident report provides a separate provider-side record supporting part of the finding.
The research appeared on November 19, followed by independent reporting
HD Moore, chief research officer at the US-based cybersecurity company Dragos, published the results of a five-month AI inference-supply-chain experiment on November 19, 2026. The team placed unique URLs and controlled credentials into roughly two million prompts submitted from July through November. Dan Goodin, a San Francisco-based senior security editor at the US publication Ars Technica, independently examined the method and findings in an article published on November 20. Those are separate dates: the primary research appeared on November 19, while the source news record was published the next day. The findings do not show that every AI provider was breached. They document specific events along sampled routes and expose the risk created when one request can pass through several organizations.
How the canaries traced exposure
Each test prompt carried a secret that had not been used elsewhere: credentials tied to empty and controlled AWS accounts, unique domains, private repository references, or URLs monitored only by the researchers. The same secret was not sent through several providers, allowing later connections or credential use to be associated with a particular route. Dragos says bait from prompts handled by at least three hosting providers was abused and that some keys were tested in tens of thousands of AWS API calls. That does not prove that every real customer prompt was stolen; canaries are a sampling mechanism. However, the later use of secrets from rejected requests indicates that, in some incidents, data was visible at an intake layer before a model completed inference.
What Phala's incident report independently records
A preliminary report published by the Phala CERT Team on November 10 says an attacker took control of an inference virtual machine on a newly deployed GPU server and could see request contents during processing. Phala advises users to treat content routed through OpenRouter to its DeepSeek V4 Flash 0731 deployment between October 21 at 00:00 UTC and November 8 at 05:35 UTC as exposed, and to rotate any credentials included in those prompts. It also says only canary credentials from requests sent on October 21 and 22 were observed being misused. This provider-side record supports the claim that the canary method detected a concrete incident. Dragos did not name all three affected providers, so the entire set of reported events should not be attributed to Phala.
Why zero data retention is not an end-to-end security guarantee
OpenRouter's current documentation defines Zero Data Retention as a policy under which an eligible inference endpoint does not persist the prompt or response after processing. It does not prevent data from reaching the model, determine the processing country, govern an application's own logs, or automatically cover plugins and external tools. Dragos reports that some leaked canaries came from requests marked ZDR and that a prompt may traverse multiple routers and providers. These findings separate two risks: misunderstanding the contractual or configuration boundary, and an actual compromise of the processing provider. ZDR controls storage after inference; it is not a guarantee that an attacker inside a provider cannot read plaintext while a request is being handled.
Checks organizations can perform now
Teams sending sensitive prompts should inventory which application routes traffic to which gateway, reseller, provider and model. They can replace default lowest-price pools with approved endpoints, enforce ZDR, data-collection and residency rules at request level, and audit application logging separately. Anyone whose requests fall inside Phala's stated exposure window should rotate API keys, passwords and tokens that appeared in prompts and review CloudTrail or equivalent records for unauthorized calls. Data-loss-prevention rules and short-lived credentials can reduce the impact of future mistakes. The next evidence to watch is disclosure from the other providers not named by Dragos, Phala's promised full root-cause analysis, and independently measurable audits showing whether routing and privacy settings continue to be honored.
Illustrative news photo · Franck V. / Unsplash ↗
Sources and verification
- Primary sourceArs Technica — About Us ↗
- Primary sourceDragos — HD Moore research report ↗
- Primary sourceDragos — company information ↗
- Primary sourceOpenRouter — Zero Data Retention documentation ↗
- Primary sourcePhala CERT Team — preliminary incident report ↗
- Reporting / cross-checkArs Technica — Dan Goodin ↗
Comments
1 commentsReader question: What is your view or technical question about this story?
Write a comment
Perspectives on this story
Discussion examples prepared by the editor.
Interesting at first glance, but “Two million canary prompts expose leaks across the AI inference supply chain” leaves me wondering about maintaining consistency when sources or models change. Is there a concrete use case?
To address this, separate observations from assumptions about maintaining consistency when sources or models change. A measurable trial would help.
If this came up in our team review, I would ask about where human approval is required first. Which changed assumption would reverse the decision?
What gives me pause here is real task costs of smaller versus larger models. A comparison that includes counterexamples would be more convincing than one success story.
Editoryal soru / Editorial question: Kurumunuz yapay zekâ istemlerinin geçtiği tüm yönlendirici ve sağlayıcıları görebiliyor mu? Can your organization identify every router and provider that handles its AI prompts?