PolicyNews
2M canary prompts expose a leaky AI supply chain
Dragos researchers sent roughly two million prompts carrying unique secrets to AI providers over five months. They found breaches at three or more providers, resellers logging prompts, and Zero Data Retention flags ignored.
· 6 min read

What happened?
HD Moore, chief research officer at Dragos and creator of Metasploit, published his team's study on November 19 (Dragos). Between July and November 2026, they sent roughly two million prompts to widely used AI services, then watched where the secrets hidden inside them turned up (Ars Technica).
Every prompt carried a unique canary that had never been used anywhere else: an AWS key, a DNS name, a URL, a password, a private git repository reference. If anyone ever uses one, they must have seen the original prompt somewhere along the chain. The AWS keys were real, tied to empty accounts with full logging. Every touch could be traced back to the exact prompt, model, router and privacy setting.
Three breached providers
The most serious finding: attackers were actively stealing data from at least three hosting providers. The bait prompts were processed alongside real customer requests, so Moore says those were almost certainly copied too. The earliest incident dates to mid-July. None of the three providers knew about the breach until Dragos told them.
The largest breach involved a couple hundred AWS keys in late October. Some came from requests the provider rejected before a model ever ran, meaning the attacker was copying requests at the front door. In the other two breaches, some stolen requests were marked ZDR. As Moore puts it, a privacy flag doesn't help when the attacker is already inside.
The attackers followed a predictable path. They first checked whether each key was live, sometimes within a few hours of the request. Then they listed S3 buckets and searched Secrets Manager and Parameter Store. Finally they tried enabling Bedrock models, creating Bedrock API keys and adding their own IAM users. In total, the bait keys were used for tens of thousands of AWS API calls.
On November 1, one attacker added ARTEX, the LLM-driven pentest framework we covered in the context of the attack on the European Commission (European Commission confirms AI-driven cyberattack). Here it only made a handful of calls to confirm keys were still live. Dragos says two of its user agents, artex-agent and artex-judge-r121, appear in no public indicator list and are worth adding to detections.
Resellers that simply logged
The second finding needed no exploit at all. Some sellers on an inference marketplace listed a model at the lowest price, enabled request logging and waited. In one case, two prompts sent through a single seller had their secret URLs opened from a home broadband connection about a day and a half later, followed by username guessing at the login page. Prompts for the same model sent through a different seller were never touched, so the leak follows the seller, not the marketplace as a whole.
In another case, someone on a Mac on a home line in France took passwords from the prompts and tried root SSH logins, typing each attempt by hand. Dragos also found a router that was quietly forwarding part of its traffic to the marketplace. It says it stopped after being notified.
ZDR on paper only
A Chinese AI lab's systems fetched the secret URLs from more than half of the prompts sent to its models between July and October, including roughly two-thirds of requests marked "no data collection" or ZDR. The scraper typically arrived two or three days after the request. The same happened when the lab's models were reached through U.S. hosting providers: the prompts flowed back to the lab.
Large U.S. providers have issues too. One kept crawling a secret URL for more than two months after a single call to its search endpoint. The strangest case: 24 days after a prompt was sent to a Chinese provider, its secret link was posted in a Feishu chat channel and opened within two minutes from three desktops and three phones in Beijing and Shanghai. Dragos says the request had gone out through a U.S.-based provider with the ZDR flag.
Traceroute for AI inference
A router typically reports a single upstream provider, which is often just another router. So Dragos built a traceroute-like tool: it uses a callback image URL and deliberately malformed requests, identifying hops from the nested error envelopes. Of about two thousand routes traced, several hundred included a system it couldn't identify, and a handful sent the image to a company not on the advertised path. One marketplace seller quietly adds about a thousand hidden tokens to every prompt.

A traced route from a marketplace to Anthropic, including an undisclosed OpenRouter hop. Source: Dragos
What Dragos recommends
- Self-host if you can afford it; it's the only option where nobody else reads your prompts. Prices cited: a 512 GB Mac Studio from $22,299, DGX Spark at $6,950, DGX Station from about $105,000, an eight-GPU B300 server from $570,000.
- Go directly to a first-tier provider: one company, one contract, one privacy policy. The smallest exposure you can buy.
- Use routers with care: pick upstream providers yourself instead of the default pool, verify that privacy settings actually change the route, and stay away from reseller marketplaces.
- Seed your own canaries: put a fake credential in the context your agents and pipelines send and treat any use as an incident. A free Thinkst Canarytoken is enough to start.
- Enable secret filtering where offered. It helps, but plenty of sensitive data doesn't look like an API key.
How solid is this?
Dragos has not named any affected provider, and the findings have not been independently verified. The company says it is working with about a dozen providers. One breached provider has published a preliminary security notice, while the first affected provider still hasn't disclosed its incident. Stolen keys and secret URLs are still being accessed months later.
Why it matters
Moore's key line: a privacy flag only binds the first hop, and AI agents will happily send whatever they can read, including credentials. An agent's context almost always holds a key, a token or an internal URL, which turns the router's "lowest price" into a security decision. OpenAI's post yesterday showed how valuable model outputs are from the vendor side (OpenAI on protecting reasoning from distillation). The Dragos study shows the other side: user prompts are just as valuable a target.
Key takeaways
- About two million canary prompts, July to November 2026.
- At least three AI hosting providers breached, none of them aware.
- Marketplace sellers harvested prompts with plain request logging.
- Several providers ignored ZDR flags.
- Advice: fewer hops, your own canaries, secret filtering.
Produced with the help of AI, reviewed by an editor. How we make our articles
Reviewed and edited by the Neural Daily desk. About


