Skip to main content
All incident reports
Incident report

OpenRouter drops three AI providers after Dragos finds months of prompt theft

Dragos seeded about two million prompts with canary links and credentials and found that three AI inference providers had been copying customer requests, some marked as never to be stored. OpenRouter has removed the providers and taken 125 endpoints out of Zero Data Retention eligibility.

· First published

The short version

  • The theft. Dragos planted unique links and credentials in about two million prompts between July and November and watched for anyone using them. It found attackers had stolen data from inside three AI inference providers, the earliest since mid-July, and none of the providers knew until the researchers told them.
  • OpenRouter's response. The gateway removed three providers, took 125 endpoints out of Zero Data Retention eligibility and notified potentially affected customers. It says its own platform was not breached.
  • Provider disclosures. Phala says an attacker took control of an inference server and could view request data while it was processed, contained on 8 November. Together AI says stored prompt samples were downloaded after a credential was committed to its public GitHub repository.
  • Not only break-ins. Some marketplace sellers turned on request logging and collected what passed through, and an unnamed Chinese AI lab fetched secret links from more than half the seeded prompts sent to its models.

Timeline

DateEvent
June-JulyPrompt samples later downloaded at Together AI are generated by requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 models.
1 JulyA third-party researcher begins placing unique canary credentials in requests.
Mid-JulyThe earliest of the three provider breaches begins.
OctoberRailway says an intrusion it suffered was likely part of a series of attacks aimed at compromising AI inference providers.
21-22 OctCanary credentials sent to Phala on these dates are later misused.
Late OctoberAWS keys are planted in single requests during short windows; every request in those windows is stolen.
1 NovThe attacker brings in the ARTEX framework to confirm keys it already held were still live.
8 NovPhala contains the incident; its exposure window closes at 05:35 UTC.
21 NovTogether AI publishes a security notice about the downloaded prompt samples.
22 NovOpenRouter says it has removed three providers from its platform and 125 endpoints from Zero Data Retention eligibility.

How the requests were taken

Dragos seeded roughly two million prompts with unique web links and credentials between July and November. A planted secret can only be learned where the prompt carrying it was read, so any use of one shows that someone saw the request. The largest of the three provider breaches involved a couple of hundred AWS keys, each sent in exactly one request during short windows in late October. Every request in those windows was taken, including some the provider had rejected before a model ran, which the firm says points to copying at the first point of arrival rather than deeper in the inference stack.

The thief tested keys within hours, listed S3 buckets, searched Secrets Manager and Parameter Store, and tried to create Bedrock API keys and add IAM users, generating tens of thousands of AWS API calls. One attacker appears to have compromised two providers at once from a single rented server, and on 1 November used ARTEX, an AI-driven penetration testing framework, only to check that keys it already held were live. HD Moore, the firm's chief research officer, wrote: "A privacy flag doesn't help when the attacker is already inside."

A second pattern needed no break-in: sellers on an AI inference marketplace listed models cheaply, enabled request logging and collected the traffic. One, working from a Mac on home broadband in France, took passwords from two prompts and typed root SSH login attempts by hand. Route tracing across about 2,000 paths found several hundred unidentified systems and one seller that adds about 1,000 hidden tokens to every prompt. An unnamed Chinese AI lab fetched secret links from more than half of the seeded prompts sent to its models between July and October, including about two-thirds of those marked no data collection or Zero Data Retention.

What the providers say

OpenRouter says the 125 endpoints it removed cover the providers the researchers flagged plus others cut as a precaution, and that any endpoint removed will need assurances and validation before its Zero Data Retention status returns. The gateway is adapting the testing approach so the checks can run continuously and is verifying network paths to find where requests and responses may be forwarded to third parties.

Phala says an attacker took control of an inference server and could view request data while it was being processed. It contained the incident on 8 November and restored the model service with added protections. Requests to the DeepSeek V4 Flash 0731 model that reached Phala through the gateway between 21 October at 00:00 UTC and 8 November at 05:35 UTC should be treated as exposed, and every secret inside them rotated. Canaries sent on 21 and 22 October were the only ones later misused, but because the tests sampled traffic rather than monitoring every request, Phala keeps the wider window.

Together AI says an unauthorised party downloaded stored samples of customer prompts after an AI coding agent committed a credential to the company's public GitHub repository. The samples came from requests to three models in June and July. The party searched them for credentials to other services and tested what it found, including five of the researchers' canaries. Together AI says it has no reason to believe the data went further, understands it has been deleted, and has made zero data retention the default for new accounts. The source also names Novita as another provider, but its text ends there.

Exposure continues

The researchers are working with about a dozen providers, and planted keys and secret links were still being accessed months after the first thefts as of this week. OpenRouter said not every affected provider has posted publicly; three have posted. Railway said in October that an intrusion it suffered was likely part of a series of attacks built to target platform companies with the end goal of compromising AI inference providers.

What is not known yet

  • Who carried out the thefts at the three providers; the research does not name the attacker.
  • How many customers, and how many genuine requests, were affected beyond the seeded canaries.
  • Which affected providers have not posted publicly.
  • The identity of the Chinese AI lab that the research declines to name.
  • What Novita disclosed about its own exposure.

What to do

If you sent data to these providers

  1. Rotate every key, token, password or other secret that appeared in prompts sent to affected models.
  2. Treat requests to DeepSeek V4 Flash 0731 routed through Phala between 21 October 00:00 UTC and 8 November 05:35 UTC as exposed.
  3. Review AWS activity for S3 bucket listings, Secrets Manager and Parameter Store searches, Bedrock API key creation and new IAM users.
  4. Stop treating a Zero Data Retention flag as protection for anything placed in a prompt.

If you run an inference service or gateway

  1. Verify network paths to find where requests and responses are forwarded to third parties.
  2. Review whether request logging is enabled and who can read what it captures.
  3. Check public repositories for credentials committed by developers or coding agents.
  4. Add the user agents artex-agent and artex-judge-r121 to detections.

For everyone else

  1. Treat any credential pasted into an AI prompt as compromised and rotate it.
  2. Keep secrets out of prompts and out of any tool that may log or forward them.

Update log

  • First published.

Sources

New incident reports by email

One email a week with the reports published since the last one.

We email you about this topic only; unsubscribe any time. See our Privacy Policy.

Want us to look at your domain?

Give us one apex domain and we'll send back a written snapshot of what is already visible from outside. No call required.