Separating AI hype from reality.

Dragos Canary Test Finds Three AI Provider Breaches, Prompt Theft, and Ignored Zero-Retention Settings

November 20, 2026 · Dragos · Security

Security firm Dragos says it seeded nearly two million AI prompts with unique, trackable secrets over five months and caught three hosting providers leaking data, marketplace resellers reading customer prompts, and multiple providers ignoring Zero Data Retention flags. The providers are not named, and none of the findings have been independently confirmed.

The experiment

From July through November 2026, Dragos researchers submitted roughly two million prompts containing unique URLs, DNS names, and real AWS keys tied to empty, fully logged accounts to widely used AI services [1]. Each secret existed only in the prompt that carried it; any use of a secret indicated that someone in the inference chain had seen the prompt. Dragos also built a route-tracing system — embedding image URLs and sending deliberately malformed requests — to map the chain of routers and providers that actually handled each request [1].

Finding 1: three providers breached

Dragos says it observed attackers actively stealing data from three different hosting providers. The earliest activity dates to mid‑July; none of the three providers knew they had been compromised until Dragos notified them [1].

The largest breach involved a couple hundred AWS keys, each sent in exactly one request to a single provider within a few windows in late October. Every request sent during those windows was stolen, including some the provider rejected before any model ran — indicating the attacker was copying requests at the provider's front door [1]. In the other two breaches, some stolen requests carried Zero Data Retention (ZDR) flags. "A privacy flag doesn't help when the attacker is already inside," the report states [1].

Once attackers had the keys, they tested them within hours, enumerated S3 buckets, searched Secrets Manager and Parameter Store, and tried to create Bedrock API keys and IAM users for persistence. The bait keys were used for tens of thousands of AWS API calls [1]. Dragos says one adversary appears to have compromised two providers simultaneously: secrets sent to both were used together from a single rented server, with tooling that evolved from a Chrome‑impersonating script to Boto3 in OrbStack to the AWS CLI on Kali Linux. On November 1 the same attacker added ARTEX, an LLM‑driven penetration testing framework; two of its user agents (artex-agent and artex-judge-r121) do not appear in public indicator lists [1].

Finding 2: marketplace resellers read prompts

The second finding required no exploit. Dragos says multiple resellers on an AI inference marketplace simply listed a model at the lowest price, enabled request logging, and waited [1]. In one case, two prompts sent through a single seller's offer were opened ~36 hours later from a home broadband connection; the user began guessing usernames at the login page. Prompts for the same model routed through a different seller were untouched, suggesting the leak followed the seller, not the marketplace as a whole [1]. In another case, a Mac on a French home connection used passwords from two prompts to attempt root SSH logins by hand, then tested several AWS keys in the same session [1]. Dragos also found a router quietly forwarding some traffic to the marketplace without disclosing the hop [1].

Finding 3: privacy settings ignored

The third finding involved no credential theft. Secret URLs in prompts were scraped repeatedly, and referenced git repositories were targeted for cloning — by companies that were never supposed to see the prompts, often in direct conflict with the privacy flags on the request [1].

The clearest example involves a Chinese AI lab. Its systems fetched the secret URLs from more than half of the prompts sent to its models between July and October, including roughly two‑thirds of requests marked "no data collection" or ZDR. The scraper typically appeared two to three days after the request [1]. Several U.S. hosting providers resell that lab's models, and prompts sent to them triggered the same scraper — meaning data flowed back to the lab despite the customer choosing a U.S. provider to keep data in the U.S. [1].

U.S. providers showed similar behavior: one continued crawling a secret URL more than two months after a single inference call to its search endpoint; another runs a scraper that returns days after each request regardless of router [1]. The strangest incident: 24 days after a prompt went to a Chinese provider, the secret link was posted into a Feishu (Lark) chat channel and opened within two minutes from three desktops and three phones in Beijing and Shanghai. Dragos had disclosed only the first eight characters of the link to the provider, so whoever shared it was reading a stored copy of the full prompt more than three weeks after a ZDR‑flagged request [1].

What is verified and what is not

Dragos's own post describes the experiment design, the three categories of findings, and that it is working with about a dozen providers; one provider has published a preliminary security notice, and one router told Dragos it stopped forwarding traffic to the marketplace [2]. However:

  • The three breached providers are not named.
  • No independent confirmation of the breaches or the scraping behavior has been found in public sources [2].
  • The share of prompts affected beyond Dragos's own samples is unknown.
  • Whether the ZDR violations were intentional or due to bugs is not established; Dragos says providers "either intentionally or due to bugs" ignore settings [1].
  • Dragos sells security products and notes this work resembles its runZero offering, a commercial interest in prompt‑leak detection [2].

Provider responses and recommendations

Dragos began notifying providers in mid‑July for the first incident and in early November for the less severe issues. Reactions have been mixed: the router forwarding to the marketplace said it stopped; several providers started investigations; but stolen keys and secret URLs continue to be accessed months later [1].

Dragos recommends: self‑host for the most sensitive work (hardware costs from $6,950 for an NVIDIA DGX Spark to $570,000+ for an eight‑GPU B300 server); go directly to a first‑tier provider; use routers carefully by choosing upstream providers yourself and verifying privacy settings actually change routing; seed your own canaries (a free Thinkst Canarytoken is enough to start); and enable secret filtering where available [1].

Related research

A separate, complementary experiment by Lauren Proehl at THOR Collective Dispatch ran a credential‑bait server from July to November 2026, capturing 155 stolen AWS keys and 18,700 SMTP sessions. It found a median 37 minutes from theft to first use, a credential economy centered on messaging (SES/SNS) and LLMjacking (Bedrock), and a clear separation between harvesters (datacenter IPs) and users (U.S. residential broadband) [7]. That work reinforces the speed and structure of the credential‑theft pipeline Dragos observed.

Sources

  1. Three breaches, endless leaks: Your AI supply chain needs canaries (dragos.com)
  2. AI Prompt Canaries: 3 Provider Breaches, ZDR Ignored (explainx.ai)
  3. Krillion Answers Today #128 (November 20, 2026) (wordhint.net)
  4. New AI Models (Nov 2026) Release Guide (clone-systems.com)
  5. AI News Today, November 20: Top Stories | AI Weekly (aiweekly.co)
  6. AI Supply Chain Security: What Canary Tokens Found | Dragos - Dragos (dragos.com)
  7. I leaked credentials on purpose for four months. Here’s who took them, and what they did next. (dispatch.thorcollective.com)