Independent/Reader-funded/Infrastructure, not tokens
DeAINEWS

AI you control — open models, private inference, and the networks that run them.

Privacy & Security

HD Moore canary experiment finds three AI inference providers breached

HD Moore's Dragos team seeded 2M prompts with canary tokens: three providers breached, marketplace resellers read prompts, and ZDR flags were ignored.

DeAI is powered by Morpheus (mor.org). We cover competing providers on the same terms — see our methodology.

A small colocation room with two charcoal server racks, amber status LEDs, and a yellow canary tag clipped to the network cable running to a lone monitoring workstation — the canary-token experiment that traced stolen AI prompts. Illustration: DeAI
A small colocation room with two charcoal server racks, amber status LEDs, and a yellow canary tag clipped to the network cable running to a lone monitoring workstation — the canary-token experiment that traced stolen AI prompts. Illustration: DeAI

HD Moore's security team at Dragos seeded about two million AI prompts with trackable canary credentials and found three inference providers compromised, marketplace resellers reading customer prompts, and privacy flags ignored — the strongest evidence yet that a Zero Data Retention setting only binds the first hop.

Key facts

What happened

HD Moore, the creator of Metasploit and now Chief Research Officer at the industrial-security firm Dragos, published the experiment on November 19, 2026. The team submitted roughly two million prompts over five months (July through November 2026) to widely used AI services, each carrying a unique URL, DNS name, and AWS credential that no legitimate system would ever touch. The keys were real, tied to empty accounts with full logging, so any use anywhere pointed back to the exact prompt, model, router, and privacy flag that carried it. Tokens started firing within days, and the July tokens are still firing.

The first finding is the most serious. Attackers were actively stealing data at three hosting providers, and Dragos judged that real customer prompts processed alongside the bait were almost certainly copied too. In the largest breach, a couple hundred bait AWS keys — each sent in exactly one request, within a few short windows in late October — were all stolen. Some were stolen from requests the provider rejected before a model ever ran, which means the attacker was copying traffic at the provider's front door rather than deeper in the inference stack. The other two breaches included stolen requests marked Zero Data Retention. Once the attacker had the credentials, they tested each key, listed S3 buckets, searched Secrets Manager and Parameter Store, and tried enabling hosted models in AWS Bedrock — tens of thousands of AWS API calls in total. On November 1, one attacker added ARTEX, an LLM-driven penetration-testing framework; its user-agents artex-agent and artex-judge-r121 do not appear in any public indicator list Dragos could find. The earliest incident dates to mid-July, none of the three providers knew until notified, one of the late-October victims has published a preliminary notice, and the mid-July victim still has not disclosed while stolen keys were still driving attempted logins as of late October.

The second finding required no exploit at all. Resellers on an AI inference marketplace could view and retain customer prompts using standard logging capabilities. In one case, two prompts went through a single seller's offer; about a day and a half later, someone on a home broadband connection opened both secret URLs in a desktop browser and started guessing usernames at a login page. Prompts for the same model that went through a different seller were never touched, so the leak follows the seller, not the marketplace. In another case, an attacker on a Mac at a French home broadband line took passwords from two prompts and tried to log in as root over SSH, typing each attempt by hand. One of the bait keys had gone to a different router — which turned out to be quietly forwarding some of its traffic to the marketplace without disclosing the hop.

The third finding is about policy failures, not intrusions. A Chinese AI lab's systems fetched the secret URLs from more than half the prompts sent to its models between July and October — including roughly two-thirds of the requests marked "no data collection" or Zero Data Retention, typically two or three days after the request. Several US hosting providers resell this lab's models, and prompts sent through those US providers triggered the same scraper; a customer who picked a US provider specifically to keep data in the US did not get what they paid for. Large US providers have their own issues: one kept crawling a secret URL more than two months after a single inference call. The strangest incident: twenty-four days after a prompt was sent to a US-based provider with the ZDR flag, its secret link was posted into a Feishu chat (Lark outside China) and opened from three desktops and three phones in Beijing and Shanghai within two minutes. Dragos's disclosure included only the first eight characters of the link, so whoever shared it was reading from a stored copy of the full prompt.

Why it matters

Every hop in an inference chain can see the full prompt and response. Routers, resellers, and upstream model hosts sit between you and the GPU, and the Dragos data shows each hop is a potential leak path — through compromise, through ordinary logging, or through a scraper that treats "no data collection" as a suggestion. As the report puts it, "a privacy flag only binds the first hop."

For builders deciding where to run models, this reframes whether your inference provider logs your prompts from a privacy-policy exercise into an operational-security question. Zero Data Retention APIs are a contract, not a control: when an attacker is already inside a provider, ZDR does nothing, and when the provider ignores the flag, it does less than nothing. The canary experiment gives the first hard measurements of how often that gap shows up in practice.

The report's recommendations map onto decisions builders already face. Host it yourself where the work is sensitive enough — the only path where nobody else reads your prompts. Prefer first-tier providers under one contract over reseller marketplaces, where capacity sometimes traces to stolen keys. If you use a router, pin the upstream providers yourself instead of accepting the default pool. And seed your own canaries: a free Thinkst Canarytoken dropped into agent context turns any leak into a detected incident. NEAR AI's SGLang remote-code-execution incident showed what a compromised inference engine can do; this report shows the supply chain around the engine leaks too.

Background

The inference market has layered routers and marketplaces between customers and GPUs because they lower cost and consolidate billing. Dragos's route traces make the hidden structure visible: of ~2,000 traced routes, several hundred contained a system the route never disclosed. Our prior coverage of OpenRouter's logging and privacy tiers described what the platforms publish about themselves; the canary data measures what actually happens to prompts in that ecosystem. Zero-data-retention guarantees have been the privacy story for API inference since 2024, and the confidential-computing stack exists precisely because policy promises felt insufficient. This report is the first large-scale empirical test of the policy layer itself, and the results are not flattering: policy is enforced unevenly, sometimes not at all, and it is unenforceable against an attacker already inside the provider.

Canary tokens are not new — Thinkst's Canarytokens have been a standard tripwire for years — but applying them at the scale of two million prompts, each attributable to a specific prompt, model, router, and privacy flag, turns anecdote into measurement.

What's next

Dragos is working with about a dozen providers on notifications and says it "wouldn't be surprised to see this expand." Expect disclosure pressure on the three breached providers — one has published a preliminary notice, the mid-July victim still has not, and stolen keys were still in use as of late October. Watch for other security teams publishing their own canary results, since the method is now public. The actionable change: treat a privacy policy as a claim to be tested, not a fact to be trusted.

Questions

What did HD Moore's canary token experiment find?
Dragos seeded ~2M AI prompts with unique canary URLs and AWS credentials from July to November 2026. Three AI inference providers were actively compromised, resellers on an inference marketplace read customer prompts via standard logging, and several providers ignored privacy settings — US-submitted prompts reached Chinese AI organizations.
Does Zero Data Retention (ZDR) actually stop prompts from being stored?
Not always, according to the Dragos canary study. Some stolen requests were marked ZDR, and one ZDR-flagged prompt surfaced in a Feishu chat 24 days after it was sent to a US provider. ZDR is a policy claim by the provider, not a technical guarantee you can verify from outside.
How do canary tokens detect prompt leaks?
A canary token is a secret that exists in only one place. Dragos embedded unique URLs, DNS names, and real AWS keys (in empty, logged accounts) in ~2M prompts. Any use of those assets anywhere on the internet traces back to the exact prompt, model, router, and privacy flag that carried it.
How can builders protect prompts sent to AI inference APIs?
Dragos recommends hosting sensitive work yourself, going directly to a first-tier provider rather than a reseller marketplace, pinning your router's upstream providers, seeding your own canary tokens, and enabling secret filtering. None of these is complete on its own.

Sources

  1. Three breaches, endless leaks: Your AI supply chain needs canaries — Dragos
  2. SANS Internet Storm Center Stormcast, November 20, 2026 — SANS ISC

About DeAI

DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.

Powered by Morpheus and StrandCMS

Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more about the Morpheus Inference API →

Sponsor disclosure — not editorial

Powered by Morpheus and StrandCMS. Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more →