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 · China

Security firm Dragos seeded nearly two million AI prompts with unique, trackable secrets across multiple providers over five months and found evidence that three hosting providers were compromised, marketplace resellers were reading customer prompts through standard logging, and Zero Data Retention settings were routinely ignored — including cases where prompts sent to U.S. providers ended up at a Chinese AI lab and one seeded link appeared in an internal company chat in China.

The experiment

From July through November 2026, Dragos researchers submitted approximately two million prompts containing unique URLs, DNS names, and AWS credentials to widely used AI services. Each secret existed only in the prompt that carried it; any use of a secret proved the prompt had been seen. The team began with Thinkst Canarytokens and later built a dedicated service that generated a unique seed per prompt, derived hostnames and credentials from it, and ran the systems those credentials pointed to — DNS, web, git, SSH, databases, and mail — with full logging. Every hit could be tied back to the exact prompt, model, router, and privacy setting [1].

Three provider compromises

Dragos calls this the most serious finding. Attackers actively stole data from three different hosting providers. The earliest incident dates to mid-July; none of the three providers knew they had been breached until Dragos notified them [1].

The largest breach involved a couple hundred AWS keys, each sent in exactly one request to one provider within a few short windows in late October. Every request Dragos sent during those windows was stolen. Some stolen keys came from requests the provider rejected before a model ever ran, indicating the attacker was copying requests at the provider's front door. In the other two breaches, some stolen requests carried Zero Data Retention (ZDR) flags [1].

Once attackers had the credentials, they tested them within hours, then enumerated S3 buckets, Secrets Manager, and Parameter Store, and attempted to enable hosted models in AWS Bedrock, create Bedrock API keys, and add their own IAM users. The bait keys were used for tens of thousands of AWS API calls [1].

One adversary appears to have compromised two providers simultaneously. Secrets sent to those two providers were used together, almost immediately, from a single rented server. The attacker's tooling evolved over three weeks: a script posing as Chrome on Windows, then Boto3 inside OrbStack, then the AWS CLI on Kali Linux. On November 1, the attacker added ARTEX, an LLM-driven penetration testing framework, which made a handful of calls to confirm keys were still live. Two user agents — artex-agent and artex-judge-r121 — do not appear in public indicator lists [1].

Dragos does not name the providers. The breaches have not been independently confirmed beyond Dragos's own telemetry, and the share of affected customer prompts beyond Dragos's samples is unknown [3].

Marketplace resellers reading prompts

The second finding required no exploit. Inference marketplaces let sellers list models at the lowest price to win requests; sellers can enable request logging and see every prompt. Dragos caught multiple sellers doing exactly that [1].

In one case, two prompts sent through a single seller's offer were opened about 36 hours later from a home broadband connection; someone began guessing usernames at the login page. Prompts for the same model sent through a different seller were never touched, indicating the leak followed the seller, not the marketplace. In another, a Mac on a French home broadband line took passwords from two prompts and attempted hand-typed SSH root logins, then tested several AWS keys in a single session. One router was quietly forwarding some traffic to the marketplace — an undisclosed hop [1].

Privacy settings ignored, data flowing to Chinese lab

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

A Chinese AI lab's systems fetched the secret URLs from more than half of the prompts Dragos 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. Several U.S. hosting providers resell this lab's models; prompts sent to those U.S. providers triggered the same scraper, meaning data flowed back to the lab despite customers choosing U.S. providers to keep data in the U.S. [1].

U.S. providers had their own issues. 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 Dragos sent a prompt 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 that provider, so whoever shared it was reading a stored copy of the full prompt more than three weeks after a ZDR-flagged request sent to a U.S.-based provider [1].

Dragos states providers ignore settings "either intentionally or due to bugs in their processing flow" but has not distinguished between the two [1].

Route tracing reveals hidden hops

Dragos built a traceroute-like system for AI inference: an image URL in each request reveals every system that fetches it, and slightly broken requests expose nested error "envelopes" from each router in the chain. About 2,000 routes have been traced. Several hundred included unidentified systems; a handful sent images to companies not on the advertised path. One marketplace seller quietly added about 1,000 hidden tokens to every prompt; another fetched images through a shared consumer subscription account [1].

Provider responses and ongoing exposure

Dragos notified all affected providers. The first provider (mid-July incident) initially thought it was a misconfiguration, then a partner data-processing issue, until investigation showed internal credentials were also compromised. As of late October, that provider had not publicly disclosed the incident, and stolen data continued to drive login attempts. For the late-October breaches, one provider published a preliminary security notice. For data-sharing issues, Dragos began notifying providers in early November and is working with about a dozen. Reactions are mixed: one router said it stopped forwarding to the marketplace, several providers started investigations, but stolen keys and secret URLs continue to be accessed months later [1].

No provider has published a statement fully confirming or disputing Dragos's findings beyond the preliminary notice [3].

Related evidence: the credential supply chain

A separate four-month experiment by Lauren Proehl, published November 19, 2026, seeded a fake developer box with canary AWS keys and SMTP credentials. Of 155 stolen AWS keys that were used, 72% went straight to SES/SNS messaging checks; median time from theft to first use was 37 minutes. Thieves (datacenter IPs) and users (residential broadband, 81% U.S.) were almost entirely different populations. LLMjacking — testing stolen keys for Bedrock model access — grew from 12% of keys in August to roughly half in September and October, spread across ~20 unrelated networks. Credential checkers tested keys against fake AI endpoints with probes designed to catch honeypots, and stolen material reached Telegram link previews within days [7].

Dragos's recommended mitigations

Dragos advises: host models yourself 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 to minimize hops; use routers carefully — pick upstream providers yourself, verify privacy settings actually change routing, and avoid reseller marketplaces; seed your own canaries (a free Thinkst Canarytoken is enough to start); enable secret filtering where available, though it won't catch sensitive data that doesn't look like a credential; and remember that AI agents will send whatever they can read, including credentials [1].

Dragos notes this research effort "felt a lot like what we do at Dragos with runZero; using small bits of data to provide a comprehensive view of networks from multiple perspectives," a commercial interest in prompt-leak detection [1].

Sources

  1. Three breaches, endless leaks: Your AI supply chain needs canaries (dragos.com)
  2. Aviatrix Threat Research Center (aviatrix.ai)
  3. AI Prompt Canaries: 3 Provider Breaches, ZDR Ignored (explainx.ai)
  4. Company Directory (infosecurity-magazine.com)
  5. The Briefing | P.K. Sharma (pk-sharma.com)
  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)