← Back to all posts
News

Dragos' Two-Million Canary Prompts Expose Three Breached AI Providers

November 22, 2026 · 02:18 UTC · News
On this page 8 sections
Dragos' Two-Million Canary Prompts Expose Three Breached AI Providers

TL;DR

Dragos has a report out that anyone shipping prompts through a third-party API should read twice. Between July and November 2026, a team led by HD Moore, the researcher who built Metasploit, sent roughly two million prompts to widely used AI services. Every prompt carried a unique canary: an AWS key, a DNS name, and a URL that existed nowhere else. If any of them was ever touched, someone in the inference chain had read the prompt. The canaries caught attackers stealing requests at three hosting providers, resellers on an inference marketplace logging customer prompts for themselves, and a Chinese AI lab whose scraper fetched the secret URLs from roughly two-thirds of the requests explicitly marked Zero Data Retention. A couple hundred planted AWS keys went on to be used for tens of thousands of AWS API calls. None of the three providers knew it was breached until Dragos called.


At a glance
What two million bait prompts found
2M
seeded prompts, July through November 2026
3
providers breached, none self-detected
~200
AWS keys stolen in late-October windows
10,000s
AWS API calls made with the bait keys
2/3
of ZDR-flagged URLs fetched by one lab's scraper
2K
inference routes traced, some hops unknown

How you wiretap an inference chain

A canary token is a secret that exists in exactly one place, so the first time anyone uses it, you know which drawer was opened. Dragos started in July with Thinkst Canarytokens, then built a dedicated seeding service: every prompt got its own derived hostname, URL, and credentials, and the AWS keys were real keys tied to empty accounts with full logging. Every hit maps back to the exact prompt, model, router, and privacy flag that carried it.

A fired canary says a prompt leaked. It does not say where, because a router reports a single upstream and that upstream is often just another router. So the team built traceroute for inference. Each request carries an image URL aimed at Dragos servers, and every system that fetches it hands over an IP address, an HTTP client, and a timestamp. The second trick is sending slightly broken requests on purpose: each router wraps the error from the hop behind it in its own format, so one response can carry three or four nested envelopes, like a forwarded email where every quote level names who passed it along. Matching the layers against a signature database yields the hop list, the software, the network, and often the operator.

About two thousand routes later, several hundred had passed through a system nobody could identify. One marketplace seller silently adds about a thousand hidden tokens to every prompt it forwards. Another fetches images through a shared consumer subscription account.

Three providers, already inside

The breaches ran from mid-July onward. In the largest, attackers copied a couple hundred AWS keys during short windows in late October, and every request sent in those windows was taken, including ones the provider rejected before a model ever ran. The copy happened at the provider's front door, upstream of the inference stack. In the other two breaches, some of the stolen requests were marked ZDR. A privacy flag does nothing once the attacker is already inside.

Anatomy of the largest breach
From seeded prompt to persistent access
Seed
One key per prompt
real AWS key, empty account, full logging
Steal
Copied at the front door
even requests rejected before a model ran
Test
Live checks within hours
Oct 21 key first used the next day
Monetize
S3, secrets, Bedrock
tens of thousands of API calls in total
Persist
Fresh IAM users
one key still in use three weeks later
All three providers learned about their breach from Dragos. One has still not disclosed it publicly.

What followed was conventional AWS abuse, executed well. Keys were tested within hours, then used to list S3 buckets, search Secrets Manager and Parameter Store, enable hosted models in AWS Bedrock, mint Bedrock API keys, and add fresh IAM users to keep the door open. A key seeded on October 21 was first used the next day and was still being used three weeks later.

A single adversary appears to have hit two of the providers at once, working secrets from both off one rented server. Its tooling drifted across three weeks of watching: a script pretending to be Chrome on Windows, then Boto3 inside OrbStack, then the AWS CLI on Kali Linux. On November 1 it added ARTEX, an LLM-driven penetration testing framework, which here only re-checked keys already held. Its two user agents, artex-agent and artex-judge-r121, appear in no public indicator list Dragos could find, so add them to your detections.

The cheapest seller was reading the mail

The second finding needed no exploit. On open marketplaces the seller with the lowest price wins the request, and a seller who wants your data just lists a popular model cheap, turns on request logging, and waits. The people Dragos caught were individuals on home broadband. In one case, someone opened both secret URLs from a single seller's traffic a day and a half later and started guessing usernames at the login page. In another, an attacker on a Mac in France took passwords from two prompts and tried to log in as root over SSH, typing each attempt by hand. Prompts for the same model through a different seller were never touched: the leak follows the seller, not the marketplace.

The hops you cannot see are the worse problem. One router was forwarding traffic into the marketplace without disclosing it, and one traced route ran from a marketplace seller to Anthropic with an OpenRouter hop in the middle that the marketplace never mentioned. If a deal on inference looks too good, the margin is your prompts.

ZDR is a setting, not a wall

One Chinese AI lab's systems fetched the secret URLs from more than half the prompts sent to its models between July and October, and from roughly two-thirds of the requests marked "no data collection" or Zero Data Retention. The scraper typically arrived two or three days after the request. Several US hosting providers resell this lab's models, and prompts sent to those providers triggered the same scraper, so picking a US provider bought nothing. One large US provider kept crawling a secret URL more than two months after a single call to its search endpoint.

One lab's scraper, July to October
The privacy flag changed nothing
share of secret URLs fetchedZDR-marked requests
All requests
over half
ZDR requests
about 2/3
Share of prompts whose secret URL the lab's systems fetched, by privacy flag on the request. US providers reselling the lab's models triggered the same scraper. Source: Dragos.

Then the Feishu incident. Twenty-four days after a ZDR-flagged request to a US-based provider, the secret link from that prompt was posted into a Feishu chat channel, and within two minutes it had been opened from three desktops and three phones in Beijing and Shanghai. Dragos' disclosure to that provider contained only the first eight characters of the link, so whoever shared it was reading a stored copy of the full prompt, more than three weeks after the request, under a flag that promised no storage.

What disclosure got

Dragos is now working with about a dozen providers. One of the late-October victims has published a preliminary security notice. The first breached provider, from mid-July, has still not disclosed, and its stolen data was still producing login attempts as of late October. The router that fed the marketplace stopped after being told. Stolen keys and URLs kept being accessed anyway, months after the original incidents.

What to do Monday

The report's own framing is the right one: where your prompts go is a decision, and if you don't make it, the lowest price will make it for you. A privacy flag only binds the first hop, and agents will happily send whatever they can read, including credentials.

Self-hosting is the only option where nobody else reads the prompts, and Dragos prices it honestly. Going directly to a first-tier provider is the smallest exposure you can buy. If you use a router, pick the upstream providers yourself and verify that your privacy settings actually change where requests land, and stay away from reseller marketplaces: some of that capacity is billed to stolen keys, and a few of those services only accept cryptocurrency for a reason. Seed your own canaries by dropping a fake credential into the context your agents already send and treating any use of it as an incident; a free Canarytoken is enough to start. Secret filtering helps, but most of what is sensitive in a prompt is not shaped like an API key.

The self-hosting ladder
What it costs to keep prompts in the building
hardware price, USDfrontier tier
DGX Spark
$6,950
Mac Studio
$22,299
DGX Station
~$105K
8x B300 srv
~$570K
Prices Dragos cites for self-hosting, from an NVIDIA desktop box and a 512GB Mac Studio to an eight-GPU B300 server. Source: Dragos.

Key Takeaways

  • Dragos seeded nearly two million prompts with unique canary tokens (AWS keys, DNS names, URLs) across five months and traced about two thousand inference routes, giving the first hard measurements of what happens to prompts after they leave your environment.
  • Three hosting providers were breached; attackers copied every request in short late-October windows, including requests rejected before a model ran, and some stolen requests carried ZDR flags. None of the three detected it themselves.
  • About 200 planted AWS keys were used for tens of thousands of AWS API calls: S3 enumeration, Secrets Manager and Parameter Store searches, Bedrock model enablement, and fresh IAM users. One key stayed in use three weeks; the ARTEX agents artex-agent and artex-judge-r121 are worth adding to detections.
  • Marketplace resellers harvested customer prompts with ordinary request logging, no exploit needed; the leak follows the seller, and undisclosed hops (a router feeding a marketplace, a hidden OpenRouter leg) are common.
  • One Chinese lab's scraper fetched secret URLs from over half of all prompts and about two-thirds of ZDR-marked ones, including prompts sent to US providers reselling its models; a ZDR-flagged link surfaced in a Feishu chat 24 days after the request.
  • Practical defenses: self-host what you can (Dragos prices it from $6,950 to $570,000), go direct to first-tier providers, pin router upstreams and verify flags actually change routing, avoid reseller marketplaces, and seed your own canary credentials, treating any firing as an incident.

Sources: Dragos, Three breaches, endless leaks: your AI supply chain needs canaries (HD Moore, Nov 19, 2026), HD Moore on X, Thinkst Canarytokens, OpenRouter, AWS Bedrock

AISecurityDragosCanary TokensPrompt LeakageZero Data RetentionAI InferenceSupply Chain
CONSOLE
$