explainx.ai350K+ learn from us
TrendingAI News TodayExplainx for Business
Pricing
explainx.ai

Upskill in AI — 16 free pathways, live workshops & bootcamps, and 50+ courses from practitioners. Plus the skills, tools, and MCP servers to practice on.

follow us

follow on google

Add explainx.ai as a preferred source

corporate training

support@explainx.ai

get started

Find your pathTake Free Evaluation

community

Join the community

learn

mind: share how you thinkpathways — start freeworkshopsbootcampscoursescompare Explainxcertificationsmock testsexplainx universitycorporate traininglearn skills & mcp

discover

skillsmcp serversexplainx mcpestacktoolsmdx readeragentsllmsdesignsdictionarypeopleagi trackerfelony benchranks

company

aboutvisionmissionteaminstructorsteach on explainxpartnershipscommunityhackathonscareers

content

daily AI newsstate of AI — live resultsblogreleasespromptsgeneratorsresource libraryfor LLMsexplainx.ai kids

solutions

all solutionsdeveloper upskillingmarketing upskillingproduct manager upskillingleadership upskilling

newsletter · weekly

Get AI news, tools, and insights in your inbox.

supportcontactprivacytermsdata rightshow we create contentsubmission guidelines

© 2026 AISOLO Technologies Pvt Ltd

explainx.ai

On this page

  • TL;DR: what the canary research found
  • How did the experiment work?
  • Finding 1: three providers were breached
  • Finding 2: marketplace resellers read customer prompts
  • Finding 3: privacy flags and retention policies were ignored
  • What is confirmed and what is not?
  • How do you protect your AI prompts?
  • Related reading
← Back to blog

explainx / blog

Dragos Seeded 2 Million AI Prompts With Canaries: Three Provider Breaches and Ignored ZDR Flags

AI Security, Data Privacy, Dragos, AI Routers, Cybersecurity

Part of AI Security

HD Moore at Dragos sent nearly two million canary-laced prompts to AI services. Findings: three provider breaches, resellers logging prompts, ZDR flags ignored. What to do.

Nov 19, 2026·9 min read·Yash Thakker
add explainx.ai
go deep
Dragos Seeded 2 Million AI Prompts With Canaries: Three Provider Breaches and Ignored ZDR Flags

Attackers used fake credentials planted in AI prompts at three different hosting providers, and none of the providers knew they had been breached until they were told. That is the headline finding from research published November 19, 2026, by HD Moore, chief research officer at Dragos, in a blog post titled "Three breaches, endless leaks: Your AI supply chain needs canaries."

Over five months, July through November 2026, the Dragos team sent about two million prompts containing unique URLs and canary credentials to widely used AI services, then watched where those secrets showed up. This article summarizes what Dragos reports, what remains unverified, and what you can do about it.

TL;DR: what the canary research found

table · 3 cols
FindingWhat Dragos reportsStatus
Provider compromisesData from at least three AI providers exposed; bait AWS keys and URLs used by attackersDragos's observation; providers not named
Marketplace resellersSellers viewed and retained prompts through standard logging, with no exploitDragos's observation
Ignored privacy settingsA Chinese AI lab fetched URLs from over half of prompts, about two-thirds of those marked ZDRDragos's observation
Routing surprisesHidden hops and undisclosed routers found by route tracingDragos's tooling
Provider responsesMixed; one preliminary notice published; Dragos says it is working with about a dozen providersDragos's account
Independent confirmationNone found in the sources we checkedOpen
Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

Agents that act on your behalf need rules. See how explainx helps secure your agents →

How did the experiment work?

A canary token is a secret that only exists in one place. If anyone uses it, you know that place was exposed. Dragos began in July with Thinkst Canarytokens, putting an AWS key, a DNS name and a URL that had never existed elsewhere into each request. Tokens started firing within days, and the July tokens were still firing in November.

The team then built a dedicated service that generates a unique seed per prompt and derives hostnames, URLs and credentials from it. The service runs the systems those credentials point to, including DNS, web, git, SSH, databases and mail. The AWS keys are real keys tied to empty accounts with full logging. Each hit links back to the exact prompt, model, router and privacy setting that carried it.

Dragos also built what it calls a traceroute for AI inference. It includes an image URL that points back to its servers, so every system that fetches it reveals an IP address and client. It then sends slightly broken requests, because each router wraps the error from the hop behind it, which exposes the chain of software and companies. About two thousand routes have been traced so far, Dragos says.

Finding 1: three providers were breached

Dragos calls this the most serious finding. It saw attackers actively stealing data from three hosting providers. The earliest incident dates to mid-July. Dragos says none of the three providers knew about their breach until it notified them.

What the attackers did

The largest breach involved a couple hundred of the bait AWS keys. Each key was sent in exactly one request to one provider, within a few short windows in late October, and every request Dragos sent during those windows was stolen. Some stolen keys came from requests the provider rejected before any model ran, which Dragos reads as copying at the provider's front door.

In the other two breaches, some stolen requests carried ZDR flags. "A privacy flag doesn't help when the attacker is already inside," the post says.

Once attackers had the keys, they tested whether the keys were live, sometimes within hours. They then listed S3 buckets, searched Secrets Manager and Parameter Store, and tried to enable hosted models in AWS Bedrock, create Bedrock API keys and add their own IAM users. In total the bait keys were used for tens of thousands of AWS API calls.

One adversary appears to have compromised two of the providers at once, Dragos says. Its tooling shifted over three weeks, from a script posing as Chrome on Windows, to Boto3 in OrbStack, to the AWS CLI on Kali Linux. On November 1, the attacker added ARTEX, an LLM-driven penetration testing framework. In Dragos's case ARTEX did little: a handful of calls confirming that already-held keys were live. Two user agents it sent, artex-agent and artex-judge-r121, do not appear in any public indicator list Dragos could find, and it suggests adding them to detections. We covered ARTEX activity separately in Check Point's Hacking on Autopilot report and the Censys exposed consoles post.

Finding 2: marketplace resellers read customer prompts

The second finding needed no exploit. Marketplace sellers only had to list a model at the lowest price, turn on request logging and wait, Dragos says. Some looked like individuals on home connections, trying canary servers one login at a time.

In one case, two prompts sent through a single seller's offer were opened about a day and a half later from a home broadband line, and someone began guessing usernames at the login page. Prompts for the same model through a different seller were never touched, which Dragos reads as the leak following the seller, not the marketplace. In another, a Mac on a French home line took passwords from two prompts and tried hand-typed SSH logins as root, then tested several AWS keys. One router, it turned out, was quietly forwarding some traffic to the marketplace.

This fits Dragos's description of how routers work: a single endpoint forwards to a pool of models, often via several hops, and every hop can see the whole prompt and response. Some sellers get cheap capacity from consumer subscriptions, startup credits or, in the shadier marketplaces, stolen AWS credentials or API keys.

Finding 3: privacy flags and retention policies were ignored

The third finding involved no stolen credential. Secret URLs in prompts were scraped, repeatedly, and git repositories they referenced were targeted for cloning, by companies that were never supposed to see the prompt, Dragos says.

The clearest case involves a Chinese AI lab. Its 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 or three days after the request. Several U.S. hosting providers resell that lab's models, and prompts sent to them triggered the same scraper. A customer who chose a U.S. provider to keep data in the U.S. "didn't get what they paid for," the post says.

Dragos also describes U.S. providers with their own issues: one kept crawling a secret URL more than two months after a single inference call to its search endpoint, and another runs a scraper that returns days after each request regardless of router.

The strangest incident: 24 days after a prompt went to one Chinese provider, the secret link was posted into a Feishu 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 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. Dragos's own write-up is not fully consistent on which provider received that request, so treat the detail with care.

The figure above is Dragos's route-trace example. It marks one hop, OpenRouter, as inferred and says that hop was not disclosed in the marketplace listing. That label describes what Dragos's tooling found on one path. It is not a finding that the router itself was breached or misbehaved.

What is confirmed and what is not?

From Dragos's own post: the experiment design and duration; the three categories of findings; the examples above; that it is working with about a dozen providers; that one provider published a preliminary security notice; that one router told Dragos it stopped forwarding to the marketplace.

Not independently confirmed: the breaches themselves, since providers are not named; the share of prompts affected beyond Dragos's own samples; and whether the named behaviors (such as ignoring ZDR) were intentional or bugs. Dragos itself says providers "either intentionally or due to bugs" ignore settings. We found no provider statements disputing or confirming the findings in the sources we checked.

Dragos sells security products and says this work resembles what it does with runZero. The research is real data from its own service, but readers should recognize that the publisher has a commercial interest in prompt-leak detection.

How do you protect your AI prompts?

Dragos says none of this means you should stop using AI services, but that "where your prompts go is a decision," and if you do not make it, the lowest price will. Its checklist:

  1. Self-host for the most sensitive work. Nobody else reads the prompts, but the hardware is costly. Dragos cites prices from about $6,950 for a DGX Spark to $570,000 or more for an eight-GPU B300 server.
  2. Go directly to a first-tier provider. One company, one contract, one privacy policy. It can still leak through breaches or data-sharing deals, but the exposure is smallest.
  3. Use routers with care. Choose upstream providers yourself, verify that privacy settings change where requests go, and stay away from reseller marketplaces.
  4. Seed your own canaries. Add a fake credential to the context your agents and pipelines already send, and treat any use as an incident. A free Thinkst Canarytoken is enough to start.
  5. Enable secret filtering. It helps with known credential formats but will not catch sensitive data that is not shaped like a key.
  6. Remember that agents send what they can read. An AI agent will pass along credentials in its context, so limit what it can see. AgentBeam, the agent security platform from the explainx.ai team, stops AI agents before they take dangerous actions.

For background on how inference infrastructure fails, see our coverage of the NEAR AI cloud inference breach and the European Commission cyberattack tied to an AI pentest tool. For consumer-side retention rules, read Gemini Keep Activity and training on chats.

Related reading

  • Check Point: Hacking on Autopilot and ARTEX
  • Censys: ARTEX exposed consoles
  • NEAR AI cloud inference breach
  • European Commission cyberattack and ARTEX
  • Gemini Keep Activity and training on chats
  • Bridgewater's AI cyber note

Sources: Dragos, "Three breaches, endless leaks," HD Moore, Nov 19, 2026; Thinkst Canarytokens.

Findings are accurate as of November 19, 2026, as reported by Dragos. Providers are not named in the source.

Spotted something out of date? Let us know.
Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

  • ✓ Free
  • ✓ One email a day, no spam
  • ✓ Unsubscribe anytime
Yash Thakker

Written by

Yash Thakker

Yash is an AI expert with over 300K learners. Join his workshops →

View Yash Thakker in People in AI →

Related posts

Nov 19, 2026

Bridgewater: One-in-Three Odds of a Destabilizing AI Cyberattack, and What Its Data Shows

Bridgewater Associates published a research note saying AI is helping attackers exploit bugs faster than defenders fix them. It says it manages money on a one-in-three chance of an economically destabilizing attack within a year. We separate its stated findings from the numbers circulating on social media.

Nov 19, 2026

European Commission Cyberattack: CERT-EU Says Hackers Used the AI Tool ARTEX

Politico and Reuters report that a cyberattack disrupted European Commission staff systems for more than 10 days, and that CERT-EU found the attackers used ARTEX, an AI-driven penetration-testing tool. This post separates what is confirmed from what is still unknown, and lists the defender steps CERT-EU published.

Nov 18, 2026

AI Patch Rush Blamed for November 2026 Outages: What Congress Was Told

After FAA ground stops, bank glitches, and an NSA maintenance window, officials told Congress the common thread may be urgent AI-driven patching—not a coordinated attack.