HD Moore reports that credentials planted in AI prompts were later used by outsiders. Dragos sent approximately two million prompts carrying unique detection tokens between July and November 2026. The results expose a practical risk: every provider or reseller handling a request becomes another organization entrusted with its contents.
The November 19 report separates three findings: evidence of provider compromise, reseller access to customer prompts, and data handling inconsistent with requested privacy settings. Restrict what an agent can read, verify which organizations process its requests, and keep production credentials out of its context.
Credential use is stronger evidence than a fetched link
A canary token is a unique credential or address used to detect unexpected access. Dragos assigned distinct tokens to individual requests, recording the model, route and privacy setting. Its cloud credentials belonged to empty test accounts with logging, rather than customer production accounts.
Moore describes evidence indicating compromise at three AI providers. In the largest reported incident, outsiders used a couple hundred planted AWS keys from requests sent during short windows in late October. The account includes requests rejected before a model ran, pointing to exposure near the provider’s request-handling entry point.
The subsequent activity went beyond opening a URL. Moore reports attempts to enumerate cloud data, find secrets, enable hosted models and create additional identities. Some keys were tested within hours. These are source-reported unauthorized uses of deliberately planted credentials, not evidence that two million customers were breached.
The report separately describes marketplace sellers accessing prompts through ordinary logging. In one example, credentials from two prompts were used in manual SSH login attempts. Requests for the same model routed through a different seller produced no observed token use. That comparison supports investigating the processing route rather than attributing the exposure to model weights alone.
A privacy setting is a promise made across a chain
Routers offer one endpoint while forwarding requests to upstream providers, sometimes through further intermediaries. Moore reports tracing about two thousand routes, including unidentified systems and undisclosed processing hops. A familiar first provider therefore does not necessarily identify every organization receiving a prompt.
The report also describes unique URLs fetched days or months after inference. Some came from requests marked zero data retention, a setting requesting that prompts and responses not be stored. One case reportedly involved a complete unique link appearing in a company chat more than three weeks after submission.
URL fetching and credential abuse establish different things. A fetch may involve search, previews, safety tooling or another processing service. Its timing and origin can raise retention or routing questions, but the request’s intended behavior and the provider’s contract still matter. Credential use and attempted account creation provide more specific evidence of misuse.
The report does not name all affected providers or supply a complete public route-by-route dataset. It says one provider published a preliminary notice, without linking that notice. Do not infer that a separately reported AI vendor incident is one of these three breaches.
Limit the data before choosing the route
The reported mechanisms support several controls for developers operating agents, models that perform tasks using tools:
- Keep production secrets outside model-readable files and tool output where possible. An agent can send material it is allowed to read, even when that material was irrelevant to the task.
- Document the actual processors and permitted fallback routes. Ask about reseller access, prompt logging, onward sharing and the scope of retention promises. Choosing a model name alone answers none of those questions.
- Prefer an explicitly approved provider path over an unreviewed reseller pool for sensitive work. Direct access reduces intermediaries, while still requiring trust in that provider’s security and data handling.
- Use purpose-built detection tokens with no production privileges. Unexpected use should prompt investigation of the associated request and route. A quiet token does not prove that no one copied the prompt.
- Treat secret filtering as one layer. Source code, customer records and confidential plans can be sensitive without resembling a recognizable API key.
When a real credential is known or reasonably suspected to have reached unauthorized hands, revoke it and review its activity. Deleting a prompt or changing a future retention setting cannot undo earlier disclosure. Our coverage of stolen cloud identities explains why model-service enablement and new identities belong in that review.
Moore says Dragos notified providers and was working with about a dozen on investigations. Some reported routing changes; token accesses continued after other disclosures. The article provides no universal repaired version or verified completion date for those responses.
Caveat: Dragos sells security products and ends the article with a runZero promotion; the unnamed-provider findings remain its research account, without an independently verified customer-impact total.