Skip to content

OpenRouter Drops 3 AI Providers After Prompt Theft Found

Why Did OpenRouter Drop 3 AI Providers?

OpenRouter Drops 3 AI Providers After Prompt Theft Found

OpenRouter off-boarded three inference providers and removed 125 endpoints from Zero Data Retention eligibility on November 22, 2026. The action followed Dragos research that found evidence of attackers inside at least three AI hosting providers, with some stolen requests flagged as Zero Data Retention. OpenRouter says its own systems were not breached.

Path of an AI prompt through a router and provider Four boxes in a row show a prompt travelling from an app or agent to a router, then an inference provider, then a model. Notes mark where the privacy flag is set, that every hop can see the prompt, and where Dragos found requests copied at a provider's front door. Your appor agent Router(a gateway) Inferenceprovider Model Sets theprivacy flag Every hop cansee the prompt Largest breach:requests copiedat the front door A privacy flag only binds the first hop
Conceptual diagram by UnpanicTech showing the route a prompt can take through an AI router. It is not a screenshot of any provider system.

Key Takeaways

  • Dragos seeded nearly two million prompts with unique canary secrets from July to November 2026.
  • It found evidence of three provider compromises, the earliest in mid-July, and none of the providers knew until told.
  • Some stolen requests carried the Zero Data Retention flag, so the flag did not protect them.
  • Neither Dragos nor OpenRouter has named the breached or removed providers, so the two groups may differ.
  • Dragos says real customer prompts were almost certainly copied too, but that is an inference.
  • If a key, token or password went through an affected route, rotate it now.

This is a new stretch of the software supply chain. Your prompt crosses a router, a host and sometimes a reseller before a model reads it, and the trust problem looks familiar from malicious npm and PyPI packages.

Teams that already hunt unsanctioned AI tools should extend the inventory to routers and upstream hosts. Our guide on how to detect shadow AI in enterprise networks covers the first half of that job, and a written incident response plan shortens the scramble if a key leaked.

One attacker in the Dragos data also used an LLM-driven pentest framework, a small real case of autonomous AI attacks. If exposed prompts held customer data, see our data breach notification response plan template for the notification side.

What did OpenRouter actually do on November 22?

OpenRouter removed three inference providers from its platform, took 125 endpoints out of Zero Data Retention eligibility, and notified all potentially affected customers. It says the 125 endpoints include providers flagged by the researchers plus others cut as a precaution. OpenRouter also states that the findings showed no breach or unauthorized access of its own systems, and it will require assurances before restoring any removed endpoint.

OpenRouter is a gateway that routes customers' AI requests to a pool of third-party providers. [5]

Definition: An AI inference router is a service that accepts one API request and forwards it to one of many model hosts, often at different companies. OpenRouter is one example. The router handles billing and provider choice, but each hop in the path can see the full prompt and response.

Its announcement lists work still underway. OpenRouter is adapting the researchers' testing approach to run checks continuously, and it is verifying network paths where requests may be forwarded to third parties. [2]

OpenRouter links public notices from Novita, Phala and Together on its page. It says not every affected provider has posted one, and it notified affected users either way. [2]

I read nothing in the post that says which three providers were removed. The three notices are provider disclosures, and OpenRouter does not label them as the removed set. [2]

The Hacker News reports that Dragos has not named its three breached providers. It also reports that OpenRouter has not named the removed three or said whether the groups match. [5]

How did Dragos catch the prompt theft?

Dragos, whose research triggered the OpenRouter removals, seeded nearly two million prompts with unique fake URLs and credentials between July and November 2026, then watched for anyone using them. A canary secret exists in only one place, so any use proves that place was exposed. The AWS keys were real keys tied to empty, fully logged accounts, so every use left a record.

Source summary: the Dragos method

Dragos chief research officer HD Moore ran the study from July through November 2026. His team submitted about two million prompts, each carrying unique URLs and canary credentials that no legitimate system would ever use. Every use of a secret was tied back to the exact prompt, model, router and privacy setting that carried it. [1]

Canaries show that a prompt leaked, not where in the chain. So Dragos built something like traceroute for AI inference, using image URLs that call back and slightly broken requests that expose each router's error format. About two thousand routes have been traced so far. [1]

Several hundred of those routes included a system Dragos could not identify. One marketplace seller quietly adds about a thousand hidden tokens to every prompt. A handful of routes sent the test image to a company that was not on the advertised path at all. [1]

How a canary secret exposes a leak Four boxes show the method: seed a unique fake secret in each prompt, the provider and router handle the request, someone uses the secret elsewhere, and an alert maps the use back to the prompt, model, router and privacy flag. Seed a uniquefake secret ineach prompt Provider androuter handlethe request Someone usesthe secretelsewhere Alert maps toprompt, model,router, flag Any use proves the place that saw the prompt was exposed. Dragos used real AWS keys tied to empty, fully logged accounts.
Conceptual diagram by UnpanicTech of the canary method Dragos describes. It is not a capture from the research.

Our reading: the tripwire design explains why the findings carry weight without a CVE. Each stolen secret arrives with a timestamp, a source network and a route. It also sets the limit, since a canary proves exposure only for the requests that carried one.

Two providers have since published notices that mention the researcher's canaries, which corroborates parts of the research. Phala says canaries sent on October 21 and 22 were later misused, and Together says five tested credentials were Dragos canaries. [3] [4]

Dragos is also a vendor. Its post ends with an invitation to start a runZero trial, so weigh that interest. The logs it describes are specific enough to check, and two providers have confirmed incidents. [1]

How did attackers steal prompts from inside the providers?

Dragos says the attacker in the largest provider breach copied requests at the provider's front door, because some stolen keys came from requests rejected before a model ever ran. Every request Dragos sent during a few late October windows was taken, including a couple hundred planted AWS keys. In the other two breaches, some stolen requests were marked Zero Data Retention, since a privacy flag cannot stop an attacker who is already inside.

After taking the keys, the attacker followed a routine. It tested whether each key was live, sometimes within hours, then listed S3 buckets and searched Secrets Manager and Parameter Store. It also tried to enable Bedrock models, create Bedrock API keys and add IAM users to keep access. [1]

The bait keys drove tens of thousands of AWS API calls. One key was seeded on October 21, first used the next day, and still in use three weeks later. [1]

Dragos believes one adversary was inside two providers at once. Secrets sent to both were used together, almost immediately, from a single rented server. [1]

The tooling changed over three weeks. It began as a script posing as Chrome on Windows, moved to Boto3 inside OrbStack, then switched to the AWS CLI on Kali Linux. [1]

On November 1 the attacker added ARTEX, an LLM-driven penetration testing framework. In this case it made a handful of calls to confirm that keys the attacker already held were still live. [1]

Both Kali and ARTEX showed up in the user agent that AWS records in CloudTrail. Dragos says that self-identification appeared in only a small fraction of traffic. Two ARTEX user agents, artex-agent and artex-judge-r121, appear in no public indicator list Dragos could find. [1]

Our analysis: searching CloudTrail for those two strings is cheap. It will not catch an attacker who changes the header, so treat a hit as strong evidence and a miss as meaningless.

Zero Data Retention is a promise about storage. It says nothing about who is already inside the server.

Source summary: where requests were taken

Dragos found that its planted AWS keys were each sent in exactly one request to one provider, all within a few short windows in late October. Every request sent during those windows was stolen, including some the provider had rejected before a model ran. That detail points to copying at the provider's front door, not deeper in the inference stack. [1]

In two other breaches, some stolen requests carried the Zero Data Retention flag. Dragos wrote that a privacy flag does not help when the attacker is already inside. The earliest incident dates to mid-July, and none of the three providers knew about its breach until Dragos notified them. [1]

Dragos adds that the bait prompts were processed alongside real customer requests, which were almost certainly copied too. That is Dragos's inference, and the report does not show what happened to real customer data. [1]

The mid-July provider first blamed an internal misconfiguration, then a partner's data processing. Its own investigation later found attackers had compromised its internal credentials. Dragos says it had still not disclosed the incident publicly. [1]

Order of events from mid-July to November 22, 2026 A vertical list of nine events in order: earliest provider breach in mid-July, Phala exposure window opens October 21, AWS key windows stolen in late October, ARTEX appears November 1, Phala contains its incident November 8, Phala report November 10, Dragos research November 19, Together AI notice November 21, OpenRouter announcement November 22. Mid JulyEarliest provider breach (Dragos) Oct 21Phala exposure window opens Late OctAWS key windows stolen at one provider Nov 1ARTEX added to attacker tooling Nov 8Phala contains its incident Nov 10Phala publishes preliminary report Nov 19Dragos publishes the research Nov 21Together AI publishes its notice Nov 22OpenRouter announces the removals Attacker activity Disclosure or response
Conceptual timeline by UnpanicTech built from the Dragos, Phala, Together AI and OpenRouter pages. Events are listed in order and are not drawn to scale.

Who else read prompts without breaking in?

Dragos found two sources of exposure that needed no intrusion. Marketplace resellers listed models at the lowest price, switched on request logging and read what flowed through. Separately, a 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 as no data collection or Zero Data Retention.

Dragos says the resellers looked like individuals on home connections. One opened both secret URLs from a desktop browser about a day and a half after the prompts and started guessing usernames. [1]

In another case, a Mac on home broadband in France took passwords from two prompts and tried root SSH logins, typing each attempt by hand. The same Mac later tested several planted AWS keys. [1]

The leak followed the seller, not the marketplace as a whole. Prompts for the same model sent through a different seller were never touched. [1]

U.S. hosts that resell the Chinese lab's models triggered the same scraper, so prompts sent to them flowed back to the lab. A git client on another cloud later tried to clone private repositories referenced in the prompts, but only for that lab's models. [1]

One secret link appeared in a Feishu chat 24 days after the request, and was opened within two minutes from three desktops and three phones in Beijing and Shanghai. Dragos had disclosed only the first eight characters of that link, so whoever shared it was reading a stored copy of the full prompt. [1]

Dragos does not name the lab. I did not find a primary source that does, so this article does not either.

Every router hop sees the whole prompt, so every hop belongs in your threat model.

What have the providers themselves disclosed?

OpenRouter's announcement links notices from three providers, and they describe different failures. Phala says an attacker took control of an inference virtual machine and could see in-flight requests. Together AI says an AI coding agent committed a credential that let an unauthorized party download stored prompt samples. Novita, as The Hacker News reports its notice, says a routing configuration issue sent some MiniMax requests onward to MiniMax's official API.

This table lays out what Phala, Together AI and Novita each disclosed alongside the OpenRouter provider removals.

Provider What it says happened Data and window at risk
Phala An attacker chained a legacy internet-facing service with management interfaces that trusted loopback access, then took over an inference VM on a new GPU server serving production traffic without full hardening. [4] DeepSeek V4 Flash 0731 requests routed by OpenRouter, from Oct 21 00:00 to Nov 8 05:35 UTC. Treat contents as exposed. [4]
Together AI An AI coding agent committed a credential to its public xorl repository. An unauthorized party used it to download stored prompt samples from a private Hugging Face repository. [3] Samples from MiniMax M3, GLM-5.2 and Kimi K2.6, mainly June 24, July 4, July 8 and July 14, 2026. [3]
Novita A routing configuration issue sent some MiniMax requests onward to MiniMax's official API, per The Hacker News. [5] Not quantified in the sources I opened. I did not open Novita's own post.

Phala says the researcher sent unique canary credentials regularly from July 1 through November 7, and only those sent on October 21 and 22 were later misused. Phala keeps a wider exposure window because the tests sampled requests. It also confirms misuse of some stolen credentials belonging to one confidential VM tenant. [4]

Phala says a full platform audit found no evidence of compromise in other models or products, and a root cause analysis will follow. Our observation: the report does not say whether the compromised server ran inside confidential computing hardware. Customers buying confidential inference will want that answered. [4]

Together says its automated scanning detected the exposed credential and the commit was reverted. Five of the credentials the unauthorized party tested were Dragos canaries. Together has no reason to believe the data went further, understands it has been deleted, and made zero data retention the default for new accounts. [3]

Our reading: Together's event involved stored samples and a repository credential, which does not obviously match copying at a front door in late October. Dragos says one of its late October providers has published a preliminary notice, and Phala's window starts October 21, which fits. Neither company has said it is a provider Dragos describes, so we are not matching them by guesswork. [1]

What do we still not know?

The OpenRouter provider removals leave four questions open. The sources I read do not name the three providers OpenRouter removed or the three Dragos found breached, and nobody has said whether they are the same. No source quantifies how many real customer requests were read. Dragos's claim that real prompts were copied is an inference from bait. And Phala and Together describe investigations that are still in progress.

  • Which providers were removed. OpenRouter's post does not say, and Dragos does not name its three either. [2]
  • How many users and requests were affected. Phala's report gives no count, and Together is still reviewing the affected prompt contents. [4]
  • Whether real prompt contents were misused. Phala confirms misuse of stolen credentials for one tenant, but no source shows what happened to prompt text itself.
  • Which provider disclosure maps to which Dragos finding.

UnpanicTech has not reproduced the canary experiment and has not tested the detection suggestions in this article.

This table separates what is confirmed from what is inference in the OpenRouter provider removals story.

Claim Status
OpenRouter removed three providers and 125 ZDR endpoints Confirmed by OpenRouter [2]
OpenRouter's own systems were breached OpenRouter says no. Vendor statement, not independently verified. [2]
Attackers were inside at least three providers Reported by Dragos. Phala and Together each published incident notices mentioning researcher canaries. [1] [3] [4]
ZDR-flagged requests were stolen Reported by Dragos, not independently confirmed [1]
Real customer prompts were copied Dragos inference, not demonstrated [1]
Prompts reached Chinese AI organizations through U.S. providers Reported by Dragos, lab not named [1]
Removed providers match the breached providers Unknown. Neither side has said.

What should teams using AI routers do now?

After the OpenRouter provider removals, rotate any secret that appeared in prompts sent through the affected routes, then decide where prompts should go instead of accepting the cheapest default. Dragos recommends self hosting for the most sensitive work, calling a first tier provider directly, choosing router upstreams yourself, avoiding reseller marketplaces, seeding your own canaries and enabling secret filtering. No single step is complete protection.

Source summary: Dragos and Phala guidance

Dragos says none of this means teams should stop using AI services. Where prompts go is a decision, and without one the lowest price makes it. A privacy flag only binds the first hop, and AI agents will send whatever they can read, including credentials. [1]

Its five recommendations are to self host for the most sensitive work, call a first tier provider directly, choose router upstreams yourself and avoid reseller marketplaces, seed your own canaries, and enable secret filtering. Secret filtering helps but misses sensitive data that is not shaped like a key. [1]

Phala's own guidance is narrower. For DeepSeek V4 Flash 0731 requests that OpenRouter routed to Phala, it tells users to rotate every key, token and password in requests sent between October 21 at 00:00 UTC and November 8 at 05:35 UTC, then review those accounts for unauthorized activity. [4]

  1. Check whether your traffic used the DeepSeek V4 Flash 0731 route through Phala in that window. If it did, treat everything in those requests as exposed. [4]
  2. Check whether you sent MiniMax M3, GLM-5.2 or Kimi K2.6 requests to Together serverless inference in June or July. Together says it will contact affected customers directly as needed. [3]
  3. Look for an OpenRouter notification. It says it notified all potentially affected customers. [2]
  4. In AWS, look for the pattern Dragos saw after key theft: unexpected key use, S3 listing, Secrets Manager and Parameter Store reads, Bedrock key creation and new IAM users. Search CloudTrail user agents for artex-agent and artex-judge-r121. [1]
  5. Seed a fake credential in the context your agents already send, and treat any use of it as an incident. [1]
Defender workflow after the OpenRouter provider removals Five stacked steps: map the route, rotate exposed secrets, hunt in cloud logs, seed your own canaries, and filter and limit what agents can read. 1. Map the route Which router, which upstream host, which flags 2. Rotate exposed secrets Any key, token or password sent on an affected route 3. Hunt in cloud logs Odd key use, Secrets Manager reads, new IAM users 4. Seed your own canaries A fake credential in agent context, any use is an incident 5. Filter and limit Scrub known secret formats and cut what agents can read
Conceptual workflow by UnpanicTech based on the recommendations from Dragos, Phala and Together AI. It is not an official procedure from any of them.

Why does this matter beyond OpenRouter?

The OpenRouter provider removals show how far a privacy label can sit from what happens to a prompt. A Zero Data Retention flag asks every hop to behave. Dragos found hops that ignored it, hops that logged anyway and hops that were breached while the flag was set.

For OpenRouter customers, provider choice is now a security decision. Dragos warns that a customer who does not choose hands the choice to the lowest price. OpenRouter says it is verifying network paths for forwarding to third parties and will require assurances and validation before restoring removed endpoints.

Agents raise the stakes. They paste whatever they can read, so one copied request can hand over keys, source code and customer data together. Limiting what an agent can read removes risk that no contract clause can.

Expect buyers to start asking for evidence instead of attestations. A provider that can show a clean canary run is more believable than one that returns a questionnaire. Until that is normal, list every secret your agents can read, because some of them will reach a router.

If an agent can read a secret, a prompt can carry it.

Sources & References

  1. Dragos, "Three breaches, endless leaks: Your AI supply chain needs canaries," HD Moore, November 19, 2026
  2. OpenRouter, "Data Processing Incidents on Inference Providers," November 22, 2026
  3. Together AI, "Security event: Prompt leakage for requests in June and July of 2026," November 21, 2026
  4. Phala, "Phala November Preliminary Incident Report," November 10, 2026
  5. The Hacker News, "OpenRouter Drops 3 Providers After Researchers Catch Months of Prompt Theft at AI Hosting Firms," November 22, 2026
NK

Cybersecurity Researcher & Technical Editor

UnpanicTech publishes independent cybersecurity research and defensive security analysis. Naseem Khan serves as Technical Editor, with a focus on vulnerability analysis, defensive security, incident response, cloud security, and practical security engineering.

Leave a Comment (Authenticated Users)