Two Million Canaries: What Leaks When You Send a Prompt to an AI Provider
Nov 21, 2026
Dragos sent about two million prompts carrying unique bait URLs and AWS keys to AI providers, routers and resellers. Three hosting providers turned out to be breached without knowing it, stolen keys were in use within hours, and zero-data-retention flags did not stop prompts being read. What the study found, what a canary token is, and how to put one in your own AI traffic.
In short. Between July and November 2026, HD Moore's team at Dragos sent about two million prompts to AI services. Each prompt carried its own unique bait: URLs and credentials that no legitimate system would ever use. When bait was used, the team knew exactly which provider had received it. They found three AI hosting providers that had been breached for months without noticing, marketplace resellers reading customer prompts, and prompts sent to US providers turning up in China. Stolen AWS keys were tried within hours and went on to make tens of thousands of API calls. Assume that every hop between you and the model can read your prompts, keep secrets out of them, and put a canary in your own AI traffic so you hear about a leak before your cloud bill tells you.
On 19 November Dragos published "Three breaches, endless leaks: Your AI supply chain needs canaries". Ars Technica covered it the next day. This post covers what the study measured, what a canary is and why it worked here, and how to run the same experiment on your own traffic.
How the experiment worked
A prompt to an AI model rarely goes straight to the company that trained the model. It may pass through a router that picks a provider for you, a marketplace where third parties resell capacity, and a hosting company that runs an open-weight model on its own GPUs. Each of those hops receives the full prompt and the full response.
Moore's team wanted to know what happens to a prompt after it leaves. They created millions of unique canaries, which are URLs and credentials that would never appear in legitimate use, and embedded them in about two million prompts. Some were web addresses on servers the team controlled. Some were real AWS access keys for empty, heavily logged accounts.
Each AWS key was sent in exactly one request, to one provider. If that key is later used, there is only one place it could have come from. The prompt, the model, the router and the privacy setting that carried it are all on record.
What they found
Three providers were breached and did not know
The team found evidence that data from at least three AI providers had been exposed. All three were hosting providers. The attacks began in mid-July, and none of the three knew about its breach until Dragos told it.
In the largest breach, "a couple hundred" of the bait AWS keys were stolen in late October. Some of them came from requests the provider had rejected before any model ran. The copying happened at the front door, before the request was even accepted.
In the other two breaches, some of the stolen requests had been sent with zero data retention (ZDR) enabled, a setting that says the prompt and response must not be stored. In Moore's words: "A privacy flag doesn't help when the attacker is already inside."
Dragos does not name the providers. We have not tried to match them to the October platform breaches, and nothing in the study says they overlap.
Stolen keys were used within hours
The attackers who had the keys followed the same sequence each time:
- Check that the key works, sometimes within hours of the theft.
- List S3 buckets and search Secrets Manager and Parameter Store for more credentials.
- Turn on Bedrock models, create Bedrock API keys and add IAM users, so the access survives the original key being revoked.
Across the bait keys this came to tens of thousands of AWS API calls. One key was planted on 21 October, used on 22 October, and still in use three weeks later.
One adversary had stolen from two different providers and used keys from both
almost immediately, from a single rented server. Its tooling changed over
time: first a script pretending to be Chrome on Windows, then boto3 inside
OrbStack, then the AWS CLI on Kali Linux.
On 1 November the same actor checked
keys with ARTEX, an open-source framework that uses language models to run
penetration tests. Its requests carried the user agents artex-agent and
artex-judge-r121. Dragos found these in no public indicator list. Tools like
Kali and ARTEX identified themselves in only a small fraction of the
CloudTrail records, so do not rely on a user-agent rule alone.
Marketplace resellers could read prompts
Some AI marketplaces let third parties resell model capacity. Dragos found that multiple resellers on one such marketplace "were able to view and retain customer prompts using standard logging capabilities." About a day and a half after a prompt passed through one of those sellers, someone on a home broadband connection opened both secret URLs in it and tried to guess logins. The leak followed the individual seller. Prompts sent through other sellers on the same marketplace did not trip.
Privacy settings did not hold
Dragos reports that prompts sent to US-based providers reached Chinese AI organisations, that some URLs were fetched months after they were sent once, and that one link ended up in an internal company chat in China.
An unnamed Chinese AI lab's systems fetched the secret URLs from more than half of the prompts sent to its models between July and October. That included about two-thirds of the requests marked "no data collection" or ZDR, and prompts sent through US hosts that resell the lab's models. Moore's summary of the privacy flags: providers "often ignore these settings."
A ZDR flag is a contractual promise made by the first company you send the request to. A router that forwards your prompt to another provider may not pass the flag on, and the next provider may not honour it.
What a canary is
A canary, also called a honeytoken or canary token, is a piece of data that looks valuable and has no legitimate use. Because no real system ever touches it, any use of it is a signal. There is nothing normal to filter out, so a canary alert is almost never a false positive. The name comes from the birds miners took underground, which showed the danger before the miners could detect it themselves.
Canaries come in a few common forms:
| Kind | What it looks like | What raises the alarm |
|---|---|---|
| Cloud decoy key | A real AWS access key or GCP service account with no permissions | Any API call made with it shows up in the cloud provider's own audit log (CloudTrail, Cloud Audit Logs) |
| Bait URL | A unique link inside a document, prompt, email or config file | A request to that URL records the time, IP address and HTTP client of whoever opened it |
| DNS canary | A unique hostname | Resolving the name is enough, even if nothing connects |
| Bait file | A plausible credentials.bak or .env.production on a server | A file-integrity agent notices the file being changed or deleted |
| Decoy record | A user or customer row that no real person owns | Any login or query against it |
| Decoy host | A machine with no real job | Any connection to it |
A canary detects use. It does not detect reading. Someone who copies a prompt and never opens the link or tries the key stays invisible. In the Dragos study the attackers wanted the keys and the URLs, so they used them, and the bait fired.
Why the alarm sits outside the leaking system
None of the three breached providers detected its own breach. Their logs and alerts were inside the systems that had been compromised. A decoy AWS key reports through AWS no matter where it was stolen from. The provider that leaked it does not have to notice anything, cooperate or even still exist.
Why every canary must be unique
A single decoy key reused everywhere tells you that something leaked, not where. Dragos minted a separate secret for each request, so each alert pointed to one provider and one route. You can do the same at a smaller scale: one decoy per provider, per router or per agent is enough to tell them apart.
Why a canary must not look like one
An intruder who lists a directory and sees /root/.canary/passwords.txt
knows to leave it alone. A canary has to look like something a real system
leaves behind: an old rotated key in ~/.aws/credentials.bak, a deploy key in
a CI directory, a database URL in a forgotten .env. Keep the record of what
is a trap in your own inventory, not in the file name or its contents.
Putting a canary in your own AI traffic
The Dragos recommendation is to seed canary credentials into the context your agents and pipelines already send. A minimal version for one team:
- Mint one decoy cloud key per destination. Create an IAM user with no policies attached for each AI provider, router or agent you want to watch. Send the CloudTrail events for those users to an alert you will actually read. Free services such as Thinkst Canarytokens can mint AWS keys and URLs for you.
- Put the key where a real one would be. Add it to a system prompt, a project instructions file or an environment block that your agent already sends, in the format your real configuration uses.
- Add a unique URL beside it for anyone who reads prompts without trying keys, such as a link to an internal-looking dashboard.
- Write down where each canary went and how to remove it. An alert from a canary you forgot you planted turns into a wasted afternoon.
- Treat any trip as an incident. It means that a provider, a reseller or something on your own side handled your prompts in a way you did not agree to. Rotate the real secrets that travelled the same route and look at what else those prompts contained.
Canaries tell you a leak happened. Preventing it depends on what you send and to whom. Dragos's other recommendations:
- Keep real secrets out of prompts. Provider-side secret filters can help, but they only catch strings shaped like keys, not customer data or source code.
- Shorten the chain. Use a first-tier provider directly, with one contract and one privacy policy, or run open-weight models on hardware you control.
- If you use a router, choose the upstream providers yourself. Check that turning on a privacy setting actually changes where your requests go, and avoid reseller marketplaces for anything sensitive.
- Hunt in CloudTrail for the sequence above: a single liveness check from an unknown IP, then S3 and Secrets Manager enumeration, then Bedrock model enablement and new IAM users. Add the ARTEX user agents to your detections.
How Helix Cyber runs canaries for you
One decoy key in one prompt is an afternoon's work. Forty decoys across cloud accounts, servers and agent configurations, kept working for a year, need bookkeeping. You have to know which ones exist and where they are, whether each one still fires, who answers when one does, and how to remove them. The Helix Cyber portal does that bookkeeping in its Tripwire view, as part of a Defence engagement that runs beside your penetration tests.
You approve every trap. A decoy lives inside your estate, so planting one needs its own signed Defence authorization, separate from the authorization for testing. Each proposed trap lists its path and the command that removes it, and nothing is planted until an owner or admin approves it.
Cloud decoys are minted and wired for you. Connect a GCP project and the portal creates the decoy service account and its key. It then deploys the alert wiring: a log sink that matches the decoy's identity, a Pub/Sub topic and a push subscription to the portal. On AWS you create the IAM user yourself and the portal gives you the webhook URL for the CloudTrail, EventBridge and SNS chain shown above.
Bait files are placed by an agent that cannot be used against you. The
host agent only makes outbound calls and runs no listener. It creates a bait
file only if nothing exists at that path, checks its checksum on a schedule,
and reports any change or removal. Names come from a catalogue of ordinary
leftovers such as credentials.bak, and the portal refuses a path containing
words like canary, honey, decoy or bait.
A trap counts once it has fired. After planting, the trap is triggered on purpose and the alert has to arrive within its time budget. For a GCP decoy the portal runs the drill itself, calling the cloud with the decoy's own key and waiting up to 15 minutes for the audit event to come back. A trap that was never tested is reported as untested, and a missing alert becomes a detection-gap finding that is rated and retested like any other.
Silence pages too. Host agents check in on a cadence, and an agent that misses three check-ins raises an alert. Cloud decoys are checked every 15 minutes: a deleted key, a disabled account or a repointed log sink counts as tampering.
A trip is an incident with a record. It lands in the event ledger with its evidence and is emailed to your organisation's owners and admins. Whoever handles it acknowledges it with a note, and the investigation stays in a shared thread on the event. Secret material that appears in a trip report is masked before it is stored.
Removal is part of the job. Every trap's removal command is recorded when it is planted. Retiring a GCP decoy deletes everything the portal created in your account, so "what did you leave behind?" has an answer on the last day.
For the AI supply chain, the decoy is the same kind of object. Mint one per provider or route, put each in the instructions or environment that agent sends, and register where it went. When one fires, you know which route leaked, and a penetration test against the same estate shows what a real key in that position would have reached.
We track the AI platform breaches of the past two months on Threat Watch, including the Fireworks AI incidents and the stolen Bedrock key campaign.
Frequently asked questions
- What did the Dragos canary study find?
- Dragos sent about two million prompts with unique bait URLs and credentials to AI providers, routers and resellers between July and November 2026. It found three hosting providers that had been breached since mid-July without knowing, marketplace resellers that could read and keep customer prompts, prompts sent to US providers reaching Chinese AI organisations, and stolen AWS keys being tested within hours and used for tens of thousands of API calls.
- What is a canary token?
- A canary token, or honeytoken, is a piece of data that looks valuable but has no legitimate use, such as an AWS key with no permissions or a unique URL. Because nothing real ever uses it, any use is a reliable signal that the data it was stored with has leaked.
- Does zero data retention protect my prompts?
- Not on its own. In the Dragos study, ZDR-flagged requests were stolen from breached providers, and an unnamed Chinese AI lab fetched URLs from about two-thirds of the ZDR-flagged prompts sent to its models. A ZDR setting is a promise by the first company that receives your request. A router may forward the prompt to providers that do not honour it, and the flag does nothing against an attacker already inside.
- Were the breached providers named?
- No. Dragos notified the providers and describes them only as hosting providers, a Chinese AI lab, an inference marketplace and resellers.
- Can a canary detect someone who only reads my prompts?
- No. A canary fires when it is used: a key is tried, a URL is fetched, a hostname is resolved. Someone who copies data and never acts on it stays invisible. That is why bait URLs are worth adding next to decoy keys, because people reading prompts often follow links even when they do not try credentials.
- How does Helix Cyber help with canaries?
- The Helix Cyber portal tracks every decoy in its Tripwire view as part of a Defence engagement. You approve each trap before it is planted. GCP decoy keys and their alert wiring are created through a cloud connection, and bait files are placed by an outbound-only host agent. Each trap is fired on purpose to prove the alert arrives, is health-checked for silence and tampering, pages your owners and admins when it trips, and is removed with a command recorded when it was planted.
- What were the ARTEX user agents?
- On 1 November one of the attackers checked stolen keys with ARTEX, an open-source framework that uses language models for penetration testing. Its requests carried the user agents artex-agent and artex-judge-r121. Dragos found them in no public indicator list, and notes that tool names appeared in only a small fraction of CloudTrail records.