AI prompts marked zero data retention were still read, Dragos canaries show
Dragos says two million canary-laced prompts exposed three compromised AI providers, reseller logging and privacy flags that did not hold, with AWS keys used within hours.

Security firm Dragos says it sent about two million prompts seeded with fake credentials and tracking links to widely used AI services and found three compromised hosting providers, resellers that logged customer prompts, and privacy flags that did not hold. Techmeme highlighted Ars Technica’s coverage of the findings.
The research, published on November 19 by Dragos Chief Research Officer HD Moore, ran from July through November 2026. The company created millions of unique canaries: URLs and credentials that no legitimate system would ever use, so any later use is a signal that a prompt was read or leaked.

Three breaches and resellers that kept prompts
Dragos reports three headline findings. First, data from at least three AI providers was exposed, and seeded prompts were later used in attempts against test systems and canary AWS accounts. Second, inference marketplace resellers could view and retain customer prompts using ordinary logging features. Third, there were wider data-handling failures: prompts from a US-based provider reached Chinese AI organizations, some URLs were opened months after a single request, and one seeded link showed up in an internal company chat in China.
The earliest breach dates to mid-July. According to Dragos, none of the three providers knew about its breach until the firm notified them. The largest involved a couple hundred of Dragos’s AWS keys, each placed in a single request to one provider in late October. Some of those keys came from requests the provider had rejected before any model ran.

The attackers moved quickly. Dragos says keys were tested, sometimes within hours, followed by listing storage buckets, searching secrets stores, and trying to enable Bedrock models, create Bedrock API keys and add IAM users. Overall, the bait keys were used for tens of thousands of AWS API calls. One key seeded on October 21 was first used the next day and was still in use three weeks later.
One adversary hit two providers at once, using secrets from both almost immediately from a single rented server. Its tooling changed over about three weeks, and on November 1 it added ARTEX, an LLM-based penetration-testing framework. Dragos says the two user agents it left, artex-agent and artex-judge-r121, appear in no public indicator list the authors could find, and it advises defenders to add them to detections.

Why a zero data retention flag did not help
In two of the three breaches, some of the affected requests carried a zero data retention (ZDR) flag. Dragos puts the lesson bluntly:
“A privacy flag doesn’t help when the attacker is already inside.”
The reseller cases show a separate weakness. Dragos says a seller only has to list a model at the lowest price, switch on request logging and wait. In one case, secret URLs from two prompts sent through a single seller were opened about a day and a half later from a home broadband connection, followed by login guessing.
According to the parts of the post that could be read only in summary form, a Chinese AI lab fetched seeded URLs from a large share of prompts, including many marked as no data collection, and Dragos was working with about a dozen providers as of early November. Those details should be checked against the full post. The Techmeme headline, summarizing Ars Technica, names OpenRouter and Vercel among the routers tested; the sections of the Dragos post available for this report name OpenRouter only in a route-trace figure caption, and do not describe Vercel’s role.
A privacy flag is only a promise from the first hop
The findings fit earlier academic work. A study by researchers at UC Santa Barbara, UC San Diego and Fuzzland, as reported by VPNCentral, tested 28 paid and 400 free LLM routers; nine injected malicious code and 17 used researcher-owned AWS canary credentials after seeing them in transit. Separately, Okta and Team Cymru have reported still-valid API keys for services including Gemini, OpenAI, Groq and OpenRouter in stolen-credential datasets. Together they suggest that credentials pasted into prompts are an active target, not a theoretical one.
Once a request passes through an aggregator or reseller, each party sees plaintext, and a flag on the request does not bind a party that was breached or that logs by design. Dragos’s advice, as summarized, is to self-host where affordable, use a first-tier provider directly, choose upstream providers deliberately when using routers, avoid reseller marketplaces, and seed canary credentials in agent and pipeline context so any use is treated as an incident.
Limits apply. This is a vendor’s own research, and Dragos did not publish the names of the three breached providers in the text read for this report. The providers’ responses are only partly known; Dragos says some opened investigations and one published a preliminary security notice. Whether each provider’s ZDR terms were technically violated is not established by the source.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Daniel ReedAI Research, Safety & Ethics AnalystDaniel Reed currently works as an AI Research, Safety & Ethics Analyst at Aiholics, writing about how changes in artificial intelligence are affecting and will affect scholarship,…



