OpenRouter, the company behind the AI model routing service of the same name, announced on November 22, 2026 that it has off-boarded three inference providers, the companies that actually run the models and return responses. The move follows reports from security researchers. OpenRouter also removed 125 endpoints from Zero Data Retention (ZDR) eligibility, and says the findings showed no breach of or unauthorized access to OpenRouter itself.
OpenRouter Removed Three Providers and Narrowed ZDR
According to OpenRouter's announcement, security researchers reported this week that certain inference providers had their data compromised or showed signs that their data policies were not being upheld. The company says it worked with the researchers and the providers, and off-boarded three providers as a result. It has not named those three.
The 125 endpoints removed from ZDR eligibility include providers flagged by the researchers as well as others removed as a precaution. OpenRouter says it will require appropriate assurances and validation before restoring ZDR eligibility to any of them. It has already notified all potentially affected customers.
Going forward, the company says it is adapting the researchers' testing approach into ongoing checks and technically verifying network paths to find where requests and responses may be forwarded to third parties. The announcement also links to notices published by Novita AI, Phala, and Together AI. However, OpenRouter notes that not all affected providers have posted publicly, and it does not say whether these three are the ones it removed.
Dragos Planted Decoys to Trace Where Prompts Went
The trigger was research published by Dragos on November 19. HD Moore, the security company's Chief Research Officer, says the team sent about two million prompts containing unique URLs and decoy credentials that alert when used (canary tokens) to widely used AI services over five months, from July to November 2026, and tracked what happened after inference.
Dragos says data was stolen from at least three AI hosting providers. The earliest incident dates to mid-July. In the largest case, around 200 of the team's AWS access keys, each sent in a single request, were stolen during a few windows in late October.
The attackers tested the keys, then listed S3 buckets, searched Secrets Manager and Parameter Store, tried to enable Amazon Bedrock models, created Bedrock API keys, and added their own IAM users. Dragos says the keys were used for tens of thousands of AWS API calls.
Dragos also disclosed that, in two of the breaches, some of the stolen requests were marked ZDR, and pointed out that a privacy setting does not help once an attacker is already inside. It further reported that prompts sent to U.S.-based providers reached Chinese AI organizations, and that one of its seeded links appeared in an internal company chat in China.
Each Provider Gave a Different Cause
The three companies whose notices OpenRouter highlighted have each published their own account. Their stated causes and the affected requests are as follows.
| Provider | Published | Stated cause | Affected requests |
|---|---|---|---|
| Phala | Nov. 10 | An intrusion while the model ran on a new GPU inference server that had not been fully hardened for production | Requests routed by OpenRouter to "DeepSeek V4 Flash 0731" between 00:00 Oct. 21 and 05:35 Nov. 8 UTC (09:00 Oct. 21 to 14:35 Nov. 8 JST) |
| Together AI | Nov. 21 | An AI coding agent committed a credential to a public GitHub repository, giving access to a private Hugging Face repository holding prompt samples used for testing | A subset of requests to "MiniMax M3," "GLM-5.2," and "Kimi K2.6" (mainly June 24, July 4, July 8, and July 14) |
| Novita AI | Nov. 22 | A routing configuration issue sent some requests for MiniMax models to MiniMax's official API instead of Novita's own infrastructure | Requests between May 28 and Nov. 11 (including those via OpenRouter) |
Phala's report asks users to treat the contents of affected requests as exposed and to rotate any API keys, tokens, passwords, and other secrets they contained. Together AI disclosed that the attacker tested five canary credentials from the Dragos experiment, and outlined fixes such as making ZDR the default for new accounts. Novita AI suspended the affected models and urges users to revoke or rotate any credentials included in their requests.
Users of Routing Services Are Exposed Too
OpenRouter lets developers reach many models through a single API, while the inference itself is handled by the registered providers. In other words, a prompt sent to OpenRouter passes through the infrastructure of whichever provider serves it.
This incident shows that a breach or misconfiguration at that downstream provider can reach the users of the routing service. The common advice across the notices is to rotate any API keys, passwords, or similar secrets that were included in prompts during the affected periods. Developers who used the affected models in those windows need to check each provider's notice for the specific conditions.
Our Take
OpenRouter's decision to remove three providers, even without naming them, and to pull 125 endpoints from ZDR including precautionary removals, is a fast response. However, ZDR is a promise that depends on how each provider operates, and as Dragos showed, it offers no protection once an attacker is inside the infrastructure. It is also significant that the problems went unnoticed for months until outside researchers planted decoys.
That said, OpenRouter's plan to adopt the researchers' methods for ongoing verification looks like a step toward the routing service acting as an auditor of its providers. For users, the most practical safeguard is likely to keep secrets out of prompts in the first place.
Summary
Prompted by Dragos's research, OpenRouter removed three inference providers and pulled 125 endpoints from ZDR eligibility. Developers building AI apps on OpenRouter or its providers need to check whether prompts sent during the affected periods contained credentials. OpenRouter says it will require assurances and validation before restoring ZDR eligibility and plans to put ongoing checks in place.