Dragos Canary Tokens Expose Compromised AI Inference Providers: What AI Router Customers Should Do Now

Rescana advisory card: Dragos canary tokens show prompt data stolen from at least three AI inference providers; OpenRouter off-boards three. No CVE.

Status: Dragos says canary tokens it planted in AI prompts were stolen from at least three AI inference providers and then used against canary AWS accounts and test systems that Dragos controls. Over five months, from July through November 2026, Dragos sent about two million prompts carrying unique URLs and canary credentials to AI services and watched where they surfaced. Its report, by Chief Research Officer HD Moore, was published at 08:00 EST on November 19, 2026. Dragos does not name the three providers. It says none of them knew about the breach until Dragos told them. OpenRouter, an AI router that describes itself as the operating layer between customers and model providers, responded in a post dated November 22, 2026. It says it has off-boarded three providers, removed 125 endpoints from Zero Data Retention (ZDR) eligibility and notified all potentially affected customers. OpenRouter says the findings "did not show any breach or unauthorized access of OpenRouter." It does not name the three providers it off-boarded. Its post links to incident notices from three providers: Novita, Phala and Together AI. OpenRouter does not say whether these are among the providers it off-boarded. No CVE, CISA KEV entry or CVSS score applies. We checked all sources at 01:26 EST on November 22, 2026.

Executive Summary

Many companies now send prompts to large language models through an AI router or aggregator: one API endpoint that passes each request on to one of many hosting providers. Dragos's research shows what that means for the data in those prompts: every hop can see the full prompt and response, some hops are not disclosed, and a privacy flag set at the first hop does not bind the ones behind it. In Rescana's reading, this is a third-party and fourth-party risk problem. The data at risk is whatever your staff, your applications and your AI agents put into prompts, including source code, customer data and credentials.

What matters for planning:

  • Attackers stole prompts at three providers, according to Dragos. Dragos says it "found attackers actively stealing data from three different hosting providers." Its bait prompts were processed alongside real customer requests, and Dragos says those "were almost certainly copied too." The earliest incident dates to mid-July 2026.
  • Stolen credentials were tested within hours. Dragos says the attacker tested its canary AWS keys, in some cases within a few hours of the request, then listed S3 buckets, searched Secrets Manager and Parameter Store, and tried to enable models in Amazon Bedrock, create Bedrock API keys and add their own IAM users to keep access. The bait keys drew tens of thousands of AWS API calls.
  • ZDR did not protect the data. Dragos says some stolen requests in two of the breaches were marked ZDR, and that in a separate set of cases providers appeared to ignore privacy settings and published policies. In Dragos's words, "a privacy flag only binds the first hop."
  • Marketplace sellers read prompts without any exploit. Dragos says sellers on an AI inference marketplace only had to list a model at the lowest price, turn on request logging and wait.
  • OpenRouter has acted on its own platform. It off-boarded three providers, removed 125 endpoints from ZDR eligibility, including endpoints removed as a precaution, and notified potentially affected customers. It says not all affected providers have posted publicly.
  • Three providers have published notices, linked by OpenRouter; neither Dragos nor OpenRouter ties them to the providers they describe. Phala says request contents for one model routed to it by OpenRouter should be treated as exposed for a window in October and November. Together AI says an unauthorized party downloaded stored prompt samples from June and July. Novita says some requests for MiniMax models were forwarded to MiniMax's own API between May 28 and November 11.

If your organization, or any supplier that handles your data, sends prompts through an AI router or a third-party inference provider, check your router and provider notices today, rotate any credential that has appeared in a prompt, review cloud logs for the activity Dragos describes, and ask your AI suppliers where their prompts actually go (Rescana guidance). This advisory is a new post, not a follow-up: no live Rescana post covers this research or these incidents.

Technical Information

This section summarizes Dragos's published findings. No public proof of concept or exploit code applies to these incidents, because no source describes a software flaw; Dragos uses the term proof of concept only for its own early canary setup. This post links to no attacker tooling and gives no steps that would help anyone repeat the attacks.

How the research worked. Dragos says it seeded prompts with unique values that no legitimate system would ever use: URLs, DNS names, passwords and AWS access keys. It started in July with Thinkst Canarytokens and then built a service that derives unique hostnames, URLs and credentials for every prompt and runs the services they point to. Dragos says the AWS keys "are real keys tied to empty accounts with full logging." Any use of one of these values shows that someone who saw the original prompt acted on it, and each value ties back to the exact prompt, model, router and privacy setting that carried it. Dragos also traced routes through the chain of routers and providers. It says it has traced about two thousand routes, that several hundred included a system it could not identify, and that a handful sent its test image to a company that was not on the advertised path.

Why routers widen the exposure. Dragos says a request often passes through more than one router before it reaches the server that runs the model, and that "every hop can see the entire prompt and response." Some routers are open marketplaces where the seller with the lowest price gets the request. Dragos says that in "the shadier marketplaces" inference is often billed to stolen AWS credentials or consumer API keys.

Finding 1: three compromised providers. Dragos's account:

  • The largest breach involved a couple hundred of its AWS keys. Each key was sent in exactly one request to one provider, within a few short windows in late October, and every request it sent in those windows was stolen. Some stolen keys came from requests the provider rejected before a model ran, which Dragos reads as the attacker copying requests "at the provider's front door," not deeper in the inference stack.
  • In the other two breaches, some of the stolen requests were marked ZDR.
  • Dragos says one adversary appears to have compromised two of the providers at the same time: secrets sent to both were used together, almost immediately, from a single rented server. The attacker's tooling changed over the three weeks Dragos watched it, from a script posing as Chrome on Windows to Boto3 running inside OrbStack to the AWS CLI on Kali Linux. On November 1 the attacker added ARTEX, which Dragos describes as an LLM driven penetration testing framework; in Dragos's case ARTEX only confirmed that keys the attacker already held were still live.
  • Dragos says one stolen key, seeded on October 21, was first used the next day and was still in use three weeks later.
  • For the first incident, in mid-July, Dragos says the provider at first thought it was a misconfiguration of internal tooling, then a problem with partner data processing, until its investigation found that the attacker had also compromised its internal credentials. Dragos says, as of its post, that this provider has not publicly disclosed the incident.

Finding 2: marketplace resellers. Dragos says sellers on an AI inference marketplace could view and keep customer prompts "using standard logging capabilities, without exploiting a vulnerability or bypassing platform controls." It describes individuals on what looked like home internet connections who opened secret URLs from prompts and tried to log in to Dragos's canary servers by hand. Prompts for the same model sent through a different seller were never touched, which Dragos reads as the leak following the seller, not the marketplace as a whole. Dragos also found one router that was "quietly forwarding some of its traffic to the marketplace," a hop it calls undisclosed. Dragos says that router told it that it has stopped.

Finding 3: privacy settings and retention policies ignored. Dragos says secret URLs in its prompts were fetched repeatedly by companies that were never supposed to see the prompts, often against the privacy settings on the request. Its examples:

  • The systems of one Chinese AI lab fetched 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, usually two or three days after the request. Prompts sent to several U.S. hosting providers that resell the lab's models triggered the same scraper.
  • One large U.S. provider kept crawling a secret URL more than two months after a single inference call. Another came back for Dragos's URLs days after each request.
  • A secret link from one prompt was posted into a Feishu (Lark) chat channel 24 days after the request and opened from devices in Beijing and Shanghai. Dragos says the request had been sent with the ZDR flag.

Dragos does not name the lab, the providers or the chat's owner.

What OpenRouter says. OpenRouter's post says researchers "reported that certain inference providers had data compromised or otherwise shown evidence that their data policies were not being upheld." Its three actions are listed in the Status section. It also says it is adapting the researchers' testing approach to run on an ongoing basis, checking network paths for requests and responses that may be forwarded to third parties, and requiring assurances and validation before it restores ZDR eligibility to any removed endpoint. OpenRouter thanks Dragos by name.

What the providers say. OpenRouter's post links to three provider notices. We read each one.

  • Phala (preliminary report, published 17:10 EST on November 10, 2026) says an attacker chained a legacy internet facing service with internal management interfaces that relied on loopback access for trust and did not independently authenticate callers, created unauthorized confidential virtual machines and obtained application credentials. Using a similar access pattern, the attacker also took control of an inference virtual machine for DeepSeek V4 Flash 0731, "making in-flight request data visible during processing." Phala dates the start of that intrusion to October 21, before the virtual machine tampering on November 4 and 5, and assesses a single actor carried out both. Users whose requests for that model were routed to Phala by OpenRouter between 19:00 EST on October 20 and 00:35 EST on November 8, 2026 should treat the request contents as exposed and rotate any secrets in them. A researcher's per request canary credentials were misused only for requests sent on October 21 and 22, as Phala dates them. Phala says it contained the incident on November 8 and has observed no successful unauthorized access to its systems since, but it dates the most recent observed misuse of data taken in the attack to November 9.
  • Together AI (dated November 21, 2026) says prompt samples it kept for software testing, from requests to MiniMax M3, GLM-5.2 and Kimi K2.6 made mainly on June 24, July 4, July 8 and July 14, 2026, were stored in a private Hugging Face repository. An AI coding agent committed a credential for that repository to Together AI's public GitHub repository. Together AI says its scanning detected the credential and it reverted the commit, but an unauthorized party obtained it during the exposure window, downloaded the samples, searched them for credentials to other services and tested them. Five of the credentials tested were canary credentials from a Dragos security experiment. Together AI says it is working with an external forensic firm and law enforcement, and has made ZDR the default for new accounts.
  • Novita (dated November 22, 2026) says a routing configuration issue caused some requests for MiniMax models on Novita to be forwarded to MiniMax's official API instead of being processed on Novita's own infrastructure. It says anyone who used MiniMax models through Novita between May 28 and November 11, 2026, directly or through OpenRouter, may have had request contents accessed or retained by a third party. Novita says the researcher saw access to URLs and hostnames from test requests but no use of the AWS test keys, and that it suspended the affected models.

The three notices describe different events. Novita, Phala and Together AI are named here only because OpenRouter links their notices. Neither Dragos nor OpenRouter says which providers make up Dragos's three compromised providers, Dragos's unnamed lab and resellers, or OpenRouter's three off-boarded providers, and we do not match any notice to them.

Affected Product Versions

No software product or version is named as vulnerable. What is affected is data, according to the sources:

  • Prompts and responses that passed through the compromised providers in Dragos's study. Dragos does not name the providers.
  • Requests routed through OpenRouter to the providers and endpoints it off-boarded or removed from ZDR eligibility. OpenRouter says it has notified all potentially affected customers.
  • Phala: requests for DeepSeek V4 Flash 0731 routed to Phala by OpenRouter between 19:00 EST on October 20 and 00:35 EST on November 8, 2026. Phala also contacted an affected confidential virtual machine tenant directly.
  • Together AI: stored prompt samples from requests to MiniMax M3, GLM-5.2 and Kimi K2.6, mainly from June 24, July 4, July 8 and July 14, 2026. Together AI says it will contact affected customers directly.
  • Novita: requests for MiniMax models on Novita between May 28 and November 11, 2026, made directly or through OpenRouter.

Dragos says prompts frequently contain "confidential source code, internal communications, strategic plans, customer information, credentials, and other sensitive business data," and that AI agents "will happily send whatever they can read, including credentials." Rescana's reading is that an agent or coding assistant with access to environment files, configuration or repositories can put live secrets into prompts without anyone typing them.

Timeline (EST)

  • May 28 to November 11, 2026. Novita's MiniMax forwarding window.
  • June 24, July 4, July 8 and July 14, 2026. Main dates of the prompt samples Together AI says were downloaded.
  • July 2026. Dragos starts seeding prompts and dates the earliest provider incident to mid-July.
  • October 20, 2026, 19:00 EST, to November 8, 2026, 00:35 EST. Phala's precautionary exposure window for DeepSeek V4 Flash 0731 requests.
  • Late October 2026. Request windows of the largest breach Dragos describes. Dragos says it contacted those providers as soon as it confirmed third-party access, and that one of them has published a preliminary security notice. Dragos does not name it, and we do not identify it.
  • November 1, 2026. Dragos says the attacker added ARTEX to its tooling.
  • Early November 2026. Dragos started notifying providers about data sharing issues that did not involve credential abuse, and says it is working with about a dozen providers.
  • November 8, 2026, 00:35 EST. Phala says containment was completed.
  • November 9, 2026. Phala dates the most recent observed misuse of data taken in the attack to this day.
  • November 10, 2026, 17:10 EST. Phala publishes its preliminary incident report.
  • November 19, 2026, 08:00 EST. Dragos publishes its research.
  • November 20, 2026, 17:32 EST. Dragos's page metadata records its last modification.
  • November 21, 2026. Together AI's notice is dated.
  • November 22, 2026. OpenRouter's post and Novita's notice are dated (dates only, no times).
  • November 22, 2026, 01:26 EST. We checked every source in this post.

Dragos says stolen keys and secret URLs "continue to be accessed, even months after the initial incident."

How to Find Your Exposure

The checks below combine the sources' guidance, which we mark, with Rescana guidance. No source publishes a detection rule set for these incidents.

  1. List every path your prompts take. Inventory the AI routers, aggregators, gateways and inference providers your teams and applications use, including tools that call models on your behalf, such as coding assistants, agent frameworks and SaaS features with built in AI. For each one, record which upstream providers it can route to and what privacy setting is applied (Rescana guidance). Dragos's advice is to pick upstream providers yourself instead of accepting the default pool.
  2. Check router and provider notices. If you use OpenRouter, look for its notification; OpenRouter says it has notified all potentially affected customers, whether or not the provider has posted publicly. Check whether you sent requests in the windows Phala, Together AI and Novita give, for the models they name (Rescana guidance on the matching).
  3. Find the secrets that went into prompts. Search your prompt logs, gateway logs and agent traces, where you keep them, for credentials, tokens, connection strings, private URLs and internal hostnames. Treat any that passed through an affected path as exposed. Phala and Novita advise rotating secrets that were in request contents (Novita as a precaution). Together AI advises customers handling sensitive prompt data to turn on ZDR and says it will contact affected customers directly.
  4. Review AWS CloudTrail for the activity Dragos describes. Dragos says the attacker tested stolen keys, listed S3 buckets, searched Secrets Manager and Parameter Store, and tried to enable Bedrock models, create Bedrock API keys and add IAM users. As Rescana guidance, look for keys used from unfamiliar networks or user agents, starting with GetCallerIdentity (the usual first test of a stolen key) and followed by calls such as ListBuckets, ListSecrets, GetSecretValue, DescribeParameters, GetParameter, GetParameters, GetParametersByPath, CreateUser, CreateAccessKey, CreateLoginProfile and AttachUserPolicy. For Bedrock, look for model access requests and model calls from unexpected principals, and for CreateServiceSpecificCredential for the Bedrock service, which is how long-term Bedrock API keys are created. Short-term Bedrock API keys leave no creation event, and S3 object reads appear only if S3 data events are logged. Apply the same idea to other cloud and SaaS logs.
  5. Search for the ARTEX user agents. Dragos says two user agents sent by ARTEX, artex-agent and artex-judge-r121, did not appear in any public indicator list it could find and "are worth adding to your detections." It says Kali and ARTEX showed up in the user agent field AWS records in CloudTrail, but only in a small fraction of the traffic. A clean search does not clear you.
  6. Check internal services named in prompts. If prompts mentioned internal URLs, repositories or hosts, review their access logs for unexpected visits, clone attempts or login attempts. Dragos saw secret URLs fetched days or months after the request and login attempts against its test servers (Rescana guidance on the review).
  7. Seed your own canaries. Dragos recommends adding a fake credential to the context your agents and pipelines already send and treating any use of it as an incident. It notes a free Thinkst Canarytoken is enough to start. Rescana's reading is that this is the only check here that can show a leak in a path you cannot inspect.

Workaround and Mitigation

The sources describe compromised and misbehaving services, not a software flaw with a fix, so mitigation is about where prompts go and what they contain. The sources' advice is the starting point.

Dragos's recommendations (summarized):

  1. Host models yourself where you can afford it, which Dragos calls the only option in which nobody else reads your prompts, and usually only for the most sensitive work.
  2. Go directly to a first-tier provider: one company, one contract and one privacy policy, which Dragos says can still leak through data sharing or a breach at the provider.
  3. Use routers with care: pick upstream providers yourself, verify that privacy settings change where requests go, and stay away from reseller marketplaces, which Dragos says draw some capacity from stolen keys.
  4. Seed your own canaries, as in How to Find Your Exposure.
  5. Turn on secret filtering where a provider offers it. Dragos says this "can't hurt," but much sensitive data in prompts does not look like an API key.

Provider and router steps. Phala also says to review accounts for unauthorized activity after rotating secrets from its window. Together AI's ZDR advice is in How to Find Your Exposure. OpenRouter's documentation says customers can enforce ZDR in account privacy settings, per model group, through guardrails or per request, and that ZDR enforcement covers provider routing for inference only, not plugins and tools such as web search.

Rescana guidance beyond the sources:

  1. Keep secrets out of prompts. Give AI agents and coding assistants the narrowest access that works, keep them away from environment files and credential stores, and scrub secrets in a gateway before requests leave your network. Prefer short-lived, scoped credentials so that a leaked key expires quickly.
  2. Rotate first, investigate second. For any credential that passed through an affected path, rotate it now, then review its use. Dragos says one of its stolen keys was first used the day after it was seeded and was still in use three weeks later.
  3. Restrict routing. Use your router's controls to allow only named providers, turn off automatic fallback to other providers, enforce ZDR where it is offered (on OpenRouter, choose "ZDR only"; its documentation says self-serve accounts otherwise allow Safety Retention endpoints, which may retain prompts), and exclude marketplace or reseller capacity for sensitive workloads. Treat ZDR as a contract claim to verify, not a technical control: Dragos found ZDR requests stolen in breaches and retained by providers.
  4. Pin sensitive workloads. Send regulated or highly sensitive data only to providers you hold a direct contract with, or to models you host.
  5. Plan for undisclosed hops. Dragos found a router forwarding traffic to a marketplace without saying so, and route traces that reached companies not on the advertised path. Ask for, and periodically test, the full chain.

We will update this advisory if Dragos, OpenRouter or the providers publish more detail, if the remaining affected providers publish notices, or if Phala publishes its full root cause analysis.

Indicators of Compromise

Dragos publishes two indicator values, both user agent strings sent by the ARTEX framework in its observations:

Indicator type Value Source and context
User agent artex-agent Dragos: sent by ARTEX against Dragos's canary AWS keys; not in public indicator lists Dragos could find
User agent artex-judge-r121 Dragos: as above

Dragos's figures mask addresses and use pseudonyms, and Dragos publishes no IP addresses, domains or file hashes. Phala says it shared attack indicators with the affected tenant and the gateway; it has not published them. Together AI and Novita publish none. Generic signs that are not published indicators, including the CloudTrail list, are in How to Find Your Exposure. We will update this advisory if any source publishes more.

Third-Party Risk (TPRM) Angle

In Rescana's reading, this is a supply chain problem with more links than most vendor reviews look at: your staff may use an AI router, your SaaS suppliers may send your data to models through one, and each router may forward to providers, resellers or other routers that neither of you chose. Dragos's finding that a privacy flag "only binds the first hop" means a supplier's ZDR claim about its first provider says little about what happens after it. Rescana has no data on which suppliers used the affected providers.

The suppliers to ask first (Rescana guidance):

  • SaaS and software suppliers with AI features that process your data, especially support, sales, document and code tools.
  • Development and IT service providers whose engineers use AI coding assistants or agents with access to your repositories, cloud accounts or credentials.
  • AI platform, agent and gateway vendors that route requests on your behalf.

Send these six questions (Rescana guidance):

  1. Do you send our data, or prompts that could contain it, to large language models? Through which routers, aggregators or providers, including any they forward to?
  2. Did any of those requests go through OpenRouter, or to Phala, Together AI or Novita in the windows their notices give? Did you receive a notice from OpenRouter or from any provider, including about an off-boarded provider or an endpoint removed from ZDR eligibility?
  3. What privacy setting applies to those requests, and how have you verified that every hop honors it?
  4. Can secrets, including credentials for our environments, reach prompts? What filtering or scoping prevents it?
  5. Have you rotated any credentials that passed through an affected path, and reviewed their use? What did you find?
  6. Do you test for leaks, for example with canary values, and will you tell us if a provider in your chain reports an incident?

Ask for evidence, not reassurance: the list of providers and routers with dates, copies of any notices received, the privacy settings in force, and the rotation and log review records. "Our provider offers zero data retention" does not answer question 3.

What to do now (Rescana guidance): start with the checks in How to Find Your Exposure and the steps in Workaround and Mitigation, then send the questions above to the suppliers that handle your most sensitive data.

Need help working out which suppliers send your data through AI routers and inference providers? Our team can help you prioritize them and track their answers. Talk to an expert

Sources

Talk to an expert

Talk to Rescana about this advisory — or about how we help TPRM teams keep reviews moving.