AI prompt canaries revealed that attackers reused credentials planted in requests rejected before a model ran, according to Dragos research published November 19 by researcher HD Moore. Three hosting providers were compromised, according to Dragos. Credentials from some requests rejected before inference were subsequently reused, Dragos says.

Moore identified compromised hosts, resellers collecting prompts through logging, and providers disregarding privacy settings, Ars Technica reported November 20. For developers choosing an inference service, the distinction matters: model selection and the companies handling a request are separate security decisions.

How AI Prompt Canaries Traced Exposure

From July through November 2026, Dragos submitted approximately two million canary-seeded prompts. Each canary was tied to an individual prompt. The AWS credentials worked, but belonged to empty accounts monitored for subsequent activity. That design connected later attempts to use a secret with the request that originally carried it.

The bait credentials generated tens of thousands of AWS API calls, Ars Technica reported. Credential reuse provided evidence of exposure, including when the original request had been rejected before a model ran. That makes the handling of incoming requests relevant alongside what happens during model processing.

Dragos separately traced approximately 2,000 routes, according to its research account. Hundreds of those routes included unidentified systems. Together, the experiments raise two connected questions: whether prompt contents escape their intended use, and which organizations handle those contents.

The Seller Changed the Outcome

AI routers select inference services using criteria including price and task suitability, Ars Technica explains. That selection makes the company serving a model another part of the purchasing decision.

Dragos says some marketplace sellers accessed prompts simply by enabling standard request logging. Different sellers serving the same model produced different observed exposure outcomes. The comparison points to the seller's handling of requests as a factor that a model name alone cannot capture.

The privacy-setting findings followed another pattern. Dragos reports that an unnamed Chinese lab fetched secret URLs from roughly two-thirds of sampled requests marked Zero Data Retention or no data collection.

Exposure Signals and Provider Disclosures

The three providers have not been publicly identified, and Dragos published no overall failure rate, Sulat AI & Tech News reported. Sulat also reports that Dragos will not release its canary service, code or data, citing secrets, provider identities and the method's operational usefulness.

Dragos reports notifying about a dozen providers; one recipient of a late-October incident notification published a preliminary notice. Sulat reports that Dragos hopes affected providers will disclose incidents themselves.

Those notifications are distinct from the canary observations.

Provider Choice Extends Beyond the Model

Moore favors direct relationships with first-tier providers and explicit upstream selection over default reseller pools, Ars Technica reports. His argument is that fewer intermediaries reduce the number of organizations handling sensitive content, although direct service still carries exposure risks.

Dragos also recommends canaries and secret filtering. These address different parts of the problem: detecting exposure and keeping sensitive credentials out of requests.

For teams maintaining a source-grounded AI workflow or a human approval boundary, provider selection adds an upstream privacy decision.

The practical implication is that an inference purchase includes a data-handling chain alongside a model. In Moore's experiments, changing the seller changed the observed exposure even when the model stayed the same.

Photo: Veronica Belmont at Ars Technica party by Blake Patterson from Alexandria, VA, USA (CC BY 2.0), via Wikimedia Commons.