Ai Engineering 9 min read

Dragos seeded 2 million canary secrets into AI prompts: three breached hosts, logging resellers, and ZDR flags that didn't hold

HD Moore's five-month experiment found attackers inside at least three AI providers and resellers reading prompts despite zero-retention flags. OpenRouter has since off-boarded three providers and pulled 125 endpoints from ZDR.

On November 19, 2026, HD Moore, chief research officer at the industrial security firm Dragos, published the results of an experiment that treated the AI inference supply chain like a network to be traced. In “Three breaches, endless leaks: Your AI supply chain needs canaries”, he describes sending approximately two million prompts, each carrying unique URLs and canary credentials, to inference providers and gateway routers from July through November 2026. Every secret appeared in exactly one request, so any later use of it pointed back to a specific prompt, model, route and privacy setting.

The results were worse than a few careless logs. Dragos found evidence that attackers had compromised at least three AI providers and were copying customer prompts, that resellers in an inference marketplace were reading prompts with ordinary request logging, and that some providers did not honor privacy flags such as Zero Data Retention (ZDR). On November 21 the fallout reached customers: OpenRouter users received an email headed “Requests you sent to MiniMax models may have been read,” and on November 22 OpenRouter published an incident post saying it had off-boarded three providers and removed 125 endpoints from ZDR eligibility.

How the canary experiment worked

Moore started in July with standard Thinkst Canarytokens and saw them fire within days. He then built a custom service that derived a unique hostname, URL and AWS access key from each seed. The AWS keys belonged to empty, fully logged accounts, so every call an attacker made with them was recorded without exposing anything real.

Because each token was sent once, a hit could be tied to the exact request that carried it: which model it named, which router or marketplace it passed through, and whether it was marked ZDR or “no data collection.” Dragos also used image URLs and deliberately malformed requests as a kind of traceroute, mapping which systems actually touched a prompt. It traced about 2,000 routes and found several hundred that passed through systems it could not identify. That matters because a request sent to one brand name on a router can be served by a different company, which may forward it to yet another.

The method is simple enough that any team can copy it, which is part of the point. Nothing here required access to a provider’s internals; Moore only watched where his own secrets surfaced.

Three providers breached, and none of them knew

The most serious finding is that attackers were already inside at least three AI hosting providers. The earliest intrusion Dragos found dates to the middle of July, and none of the three providers knew about their breach until Dragos notified them, starting in early November. Dragos says it contacted about a dozen providers in total and does not name the breached ones.

In the largest breach, attackers took “a couple hundred” of the bait AWS keys, all within a few short windows in late October. Some of those keys came from requests the provider rejected before a model ever ran, which means the attacker was collecting traffic at a layer in front of inference, not reading model outputs. In the other two breaches, some stolen requests were marked ZDR. As Dragos puts it, a privacy flag does not help when the attacker is already inside.

The attackers moved fast and methodically. Keys were tested in some cases within a few hours, then used to list S3 buckets, Secrets Manager and Parameter Store entries, then to enable Bedrock models, create Bedrock API keys and add IAM users: the standard playbook for turning a leaked cloud key into free model access and persistence. Dragos counted tens of thousands of AWS API calls in total. One key seeded on October 21 was first used the next day and was still in use three weeks later. One adversary hit two providers and used secrets from both together, almost immediately, from a single rented server. Its tooling changed over those three weeks, from a script spoofing Chrome to Boto3 running in OrbStack to the AWS CLI on Kali Linux.

From November 1, some of that traffic came from ARTEX, the LLM-driven attack framework that has shown up in several recent cloud breach reports. Its user agents identified themselves as artex-agent and artex-judge-r121, and it made a handful of calls to confirm that keys the attacker already held were still live. Dragos notes that self-identifying traffic was only a small fraction of the total, so user-agent filtering would not have caught most of it.

The uncomfortable implication is in one line of the report: the canary prompts were processed alongside real customer requests, and “those were almost certainly copied too.” Real prompts routinely contain API keys, database URLs, internal hostnames and source code.

Resellers, ignored flags and the MiniMax routing

The second finding concerns inference marketplaces, where third-party sellers list access to popular models. Dragos describes a low-effort pattern: list a model at the lowest price, enable request logging, and wait. In one example, two prompts sent through a single seller had their URLs opened by a home-broadband user about a day and a half later, who then tried guessing usernames. The same model bought through a different seller was never touched. The leak followed the seller, not the model.

The third finding is that privacy settings were not reliably honored. According to Dragos, one Chinese AI lab, which it does not name, fetched secret URLs from more than half of the prompts sent to its models between July and October, including roughly two thirds of requests flagged ZDR or “no data collection,” typically two to three days after the request. US providers that resold the same lab’s models triggered the same scraper. Dragos also saw a git client on another cloud try to clone private repositories referenced in prompts, only for that lab’s traffic.

OpenRouter’s customer email, which users began posting on Hacker News on November 21, fills in names that Dragos left out. As relayed by The CyberSec Guru and explainx, it says test URLs in prompts sent to MiniMax models between July and October were opened by a third party 40 to 130 hours after the request, and that three providers, AtlasCloud, Novita and StreamLake, forwarded OpenRouter requests for MiniMax models to MiniMax itself. According to the same reports, MiniMax told StreamLake it had content moderation enabled before October 22 and ZDR since, while OpenRouter said it had asked MiniMax for its own account and had not received one. Neither OpenRouter nor Dragos has stated on the record that the email and the Dragos report describe the same findings, though the timing and details line up closely.

OpenRouter’s November 22 post says the findings “did not show any breach or unauthorized access of OpenRouter,” and lists what it has done: off-boarded three providers (not named), removed 125 endpoints from ZDR eligibility, including precautionary removals beyond those the researcher flagged, and notified all potentially affected customers. It says it is adding ongoing canary-style checks and technical verification of network paths before restoring ZDR status, and links to notices from Novita, Phala and Together AI, the last of which covers prompt leakage for requests in June and July.

Why ZDR is a contract, not a control

The core lesson is that a ZDR label on a router is a promise made by whoever is at the end of the route, and the route can be longer than it looks. A request can pass from a router to a reseller to a model’s original lab, and each hop is a separate company with its own logging, its own security posture and its own interpretation of a privacy flag. Data processing agreements give you recourse after the fact; they do not stop an intruder sitting on a provider’s ingress or a seller with logging turned on.

It also reframes prompts as a credential store. Coding agents and chat tools paste environment files, stack traces with connection strings and whole repositories into prompts every day. If a provider’s ingress is compromised, that is the same exposure as a leaked secrets vault, and the Dragos AWS logs show attackers treat it that way: verify the key within hours, enumerate, then entrench.

Dragos’s recommendations follow from that. Host models yourself, or send traffic directly to a first-tier provider rather than through a chain of intermediaries. If you use a router, pin the specific upstream providers and verify their privacy settings rather than trusting a label; on OpenRouter that means restricting routing to named providers and disabling fallbacks. Avoid reseller marketplaces for anything sensitive. Seed your own canaries into representative traffic so you find out about a leak from your own tripwire rather than from a researcher’s email. And filter secrets out of prompts before they leave your network, since a key that never reaches a provider cannot be stolen from one.

What to watch

The three breached providers are still unnamed, and so far only Novita, Phala and Together have published notices that OpenRouter links to. Expect more disclosures as Dragos’s notifications work through the roughly dozen providers it contacted, and watch whether any of them can say how attackers got in and how long prompt capture ran.

MiniMax has not given OpenRouter its own account of what happened to the forwarded prompts, and Dragos has not said whether it notified the unnamed Chinese lab. How OpenRouter verifies network paths before restoring ZDR to the 125 endpoints will be a useful template for every router and marketplace, since all of them face the same problem of promising behavior on infrastructure they do not run.

For teams, the immediate step is unglamorous: if any credential has ever been pasted into a prompt sent through a router or reseller, especially to MiniMax models between June and October, rotate it now and check your cloud logs for activity you don’t recognize.

Get Insanely Good at AI

Get Insanely Good at AI

The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.

Keep Reading

Ai Agents

How to Secure AI Agents With Google ADK

Learn how to secure your autonomous workflows and prevent unauthorized actions using Google ADK's hardware-backed tool binding and execution logs.

Ai Engineering

Hugging Face hands LLM routing to OpenRouter: failover by default, custom provider keys gone

On November 16, Hugging Face replaced the custom router behind Inference Providers with OpenRouter's for all language models. Requests now fail over across providers inside one call, pricing is unchanged, and bring-your-own provider keys stop working for chat models.

Ai Engineering

Railway's breach report: a build escape, 100,000 workspaces and $740,000 in stolen AI usage

Railway's November 10 incident report shows how an attacker chained a build-isolation zero-day, a Temporal trust flaw and an OIDC misconfiguration to decrypt customer environment variables and hunt for AI provider keys.

Ai Engineering

Fireworks AI's second breach reached customer datasets and incomplete model weights

A November 9 update from Fireworks AI says an October 24 intrusion through a third-party cloud provider exposed billing data and, for some accounts, secrets, datasets and incomplete model weights. The first bulletin had said no such access was detected.

Ai Coding

Apiiro finds real secrets in 1,122 public Claude Code settings files, and 215 of them evaded two scanners

Choosing "Yes, and don't ask again" can save a command's password verbatim in .claude/settings.local.json. Apiiro counted 2,539 distinct secrets across 1,994 public repositories.