Your AI Chats May Not Have Been Private: OpenRouter Cuts Three Providers After Researchers Catch Months of Prompt Theft
Millions of people now paste their work, their code, and sometimes their secrets into AI chatbots — trusting a promise that what they type is never stored. This week, that promise broke. OpenRouter, the gateway that routes AI requests to dozens of third-party model providers, announced on November 22 that it has removed three providers from its platform and stripped 125 endpoints of their Zero Data Retention status after researchers caught providers copying, keeping, and even losing customer prompts.
The findings did not show a breach of OpenRouter itself, the company said. But they showed something arguably worse: the privacy guarantees users relied on were only as strong as the least trustworthy company in the chain — and some of those companies were compromised for months without knowing it.
How Researchers Caught It
The investigation came from Dragos, an industrial cybersecurity company, which published its research on November 19. Between July and November 2026, Dragos researchers sent nearly two million test prompts to widely used AI services. Hidden inside those prompts were unique links and test credentials known as canary tokens — digital tripwires that reveal themselves the moment someone opens a link or tries to use a credential.
When a link was opened or a credential was used, the researchers could trace it back to the original request. What they found, according to Cyber Security News, was evidence of compromises at three hosting providers. Some test requests were copied even though the model had rejected them before processing. And some requests marked Zero Data Retention — a policy that says prompts and replies should not be stored — appeared to have been kept or forwarded anyway.
In one striking case, a test link sent to a U.S. provider surfaced in a company chat in China 24 days later.
Dragos traced about 2,000 inference routes in the course of the work, according to Unite.AI. Several hundred included an unidentified system in the path, and in a handful of cases a company absent from the advertised route fetched the test image. The researchers also reported that one unnamed Chinese AI lab’s systems fetched secret URLs from more than half of the prompts sent to its models between July and October 2026 — including roughly two-thirds of the prompts marked no-data-collection or Zero Data Retention.
The largest compromise involved a couple hundred seeded AWS keys sent during several short request windows in late October 2026. And the researchers are still working with about a dozen providers, The Hacker News reported — with planted keys and secret links still being accessed months after the first thefts, as of this week.
Three Providers, Three Different Failures
OpenRouter’s announcement links public notices from three providers — Novita, Phala, and Together AI — though neither OpenRouter nor Dragos has named the three providers that were removed, and neither has said publicly whether they are the same three.
At Phala, an attacker took control of an inference server and could view request data while it was being processed, the provider said in a preliminary incident report. The incident was contained on November 8, and the model service has since been restored with added protections. Requests to the DeepSeek V4 Flash 0731 model that reached Phala through the gateway between October 21 at 00:00 UTC and November 8 at 05:35 UTC should be treated as exposed, the report says — and every key, token, password, or other secret within them should be rotated. A third-party researcher had been placing unique canary credentials in requests since July 1, and only the ones sent on October 21 and 22 were later misused, but Phala kept the wider window because the tests sampled requests and could not reveal every possible data-access event.
At Together AI, an unauthorized party downloaded stored samples of customer prompts after an AI coding agent committed a credential to the company’s public GitHub repository, the company said in a security notice published November 21. The samples had been collected for routine software testing and were generated by requests to the MiniMax M3, GLM-5.2, and Kimi K2.6 models in June and July. The intruder searched the samples for credentials to other services and tested them — five of the tested credentials were canaries from the researchers’ experiment. Together AI said 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.
At Novita, a routing configuration issue caused some MiniMax requests to be sent on to MiniMax’s official API, beyond what was intended for those customers. Requests made directly through Novita or routed through OpenRouter between May 28, 2026, and November 11, 2026, may have had their contents accessed or retained by a third party, according to the provider’s privacy notice. Novita suspended the affected models. The researcher observed accesses to test URLs and hostnames but no use of AWS test-key markers; automated content review may have been involved, an assessment Novita is checking with MiniMax.
What OpenRouter Did — and Did Not Say
In its November 22 incident notice, OpenRouter said it has off-boarded three providers based on the results of its investigations, removed 125 endpoints from Zero Data Retention eligibility, and notified all potentially affected customers. The company stressed that the 125 removals do not represent 125 confirmed breaches: the list includes providers flagged by the researchers plus additional endpoints removed as a precaution.
OpenRouter said it is checking whether providers forward requests or replies to other companies, and plans to verify endpoints before restoring their eligibility. It also said it is grateful for the Dragos report and, as it integrates similar testing techniques into its own platform, will keep raising the bar — and providers that want to stay on OpenRouter will need to meet it.
What OpenRouter did not say is how many users or conversations may have been exposed. It also noted that not all affected providers have posted publicly, though it said it has notified affected users regardless and is encouraging all providers to publish incident write-ups.
Why This Matters Beyond OpenRouter
The OpenRouter incident exposes a structural problem with how AI is sold. When a user sends a prompt through a gateway, the company that first receives the request may not be the only one that handles it. If a provider forwards the request to another service, that service can also see the prompt — and the user has no way of knowing it happened.
That is the supply-chain risk hiding inside multi-provider model routing: a “Zero Data Retention” badge on a gateway means little if a downstream host quietly caches, logs, or routes prompts across untrusted third parties. The Dragos experiment proved the point empirically — with canary tokens that showed, in concrete terms, whose hands customer data actually passed through.
The timing makes the lesson sharper. AI assistants are now woven into everyday work: drafting contracts, debugging code, summarizing medical records, planning strategy. Every one of those prompts can contain credentials, customer data, or trade secrets. The researchers’ advice, echoed by Cyber Security News, is blunt: customers who received a notice should review what they sent through the affected provider, revoke or rotate any passwords or API keys in those prompts, and check access logs for unusual activity. For future requests, restrict sensitive workloads to approved providers — and keep secrets out of prompts rather than relying on retention settings alone.
The Bottom Line
OpenRouter did the right thing once the evidence was in front of it: it cut three providers, pulled 125 endpoints from Zero Data Retention eligibility, and told affected customers. But the Dragos research shows the deeper problem was not one bad provider. It was a system in which a privacy promise made at the front door could be quietly broken three doors down — and nobody, including the providers themselves, noticed for months. Until gateways verify every provider in the chain the way Dragos did, with planted tripwires rather than policy documents, “we don’t store your data” will remain a claim, not a guarantee. Treat every AI prompt as if someone else might read it — because this week, someone did.
Sources: Cyber Security News, “OpenRouter Users’ Private AI Chats Possibly Exposed Via Compromised Third-Party Providers,” November 22, 2026; The Hacker News, “OpenRouter Drops 3 Providers After Researchers Catch Months of Prompt Theft at AI Hosting Firms,” November 22, 2026; OpenRouter, “Data Processing Incidents on Inference Providers,” November 22, 2026; Unite.AI, “OpenRouter Cuts Three Providers After AI Prompt Exposure Findings,” November 22, 2026.
Edited by Muse @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous
Never miss a customer call again — MAJ.COM’s AI receptionist answers every call, texts customers back instantly, and books the appointment for you. See how it works at https://MAJ.COM
Discover more from QUE.com
Subscribe to get the latest posts sent to your email.
