OpenRouter drops 3 providers after Dragos caught prompt theft

Share
OpenRouter drops 3 providers after Dragos caught prompt theft

Dragos spent five months planting bait secrets inside AI prompts and watching where they turned up. Three inference providers were already compromised before the test ended, and the two largest leaks needed no exploit at all.

Dragos sent roughly two million canary prompts to AI services between July and November 2026, and found three hosting providers where attackers were actively lifting customer data, marketplace resellers reading requests just by switching on logging, and privacy settings that providers ignored, according to the firm's write-up published on November 19. OpenRouter, the router most of that traffic passed through, off-boarded three providers and stripped 125 endpoints of their Zero Data Retention (ZDR) eligibility, it said in its own notice. OpenRouter says the research found no breach of its own systems, and it has not said whether the three providers it removed are the same three Dragos found compromised.

The method is the reason to trust the findings. Every test prompt carried a unique AWS key tied to an empty, fully logged account, plus its own hostname, URL, git remote and SSH login. Nothing legitimate would ever touch them, so any hit is proof that someone downstream of the API saw the prompt. Dragos started with Thinkst Canarytokens in July, then built a service that mints a fresh set of secrets per request and correlates each hit back to the exact prompt, model, router and privacy flag that carried it. The July tokens were still firing when the research ended.

The breach was at the front door

The most serious finding involved a couple of hundred seed AWS keys. Each went out in exactly one request, all within a few short windows in late October, and every request sent in those windows was stolen. Some of the keys came from requests the provider rejected before a model ever ran, which tells Dragos the attacker was copying traffic at the provider's entry point rather than somewhere deeper in the inference stack.

The follow-through was methodical. Keys were tested for liveness within hours, sometimes the next day. The attacker listed S3 buckets, searched Secrets Manager and Parameter Store, tried enabling hosted models in AWS Bedrock, created Bedrock API keys, and added IAM users to hold access open. Dragos counted tens of thousands of AWS API calls against its bait credentials, and one key seeded on October 21 was still being used three weeks later. The tooling drifted over that window — a script pretending to be Chrome on Windows, then Boto3 inside OrbStack, then the AWS CLI on Kali Linux. On November 1 the same actor added ARTEX, an LLM-driven penetration-testing framework; two user agents it sent, artex-agent and artex-judge-r121, do not appear in any public indicator list Dragos could find.

One adversary appears to have compromised two of the providers at once. Secrets sent to both were used together, almost immediately, from a single rented server.

The discount was the attack

The second finding is worse for anyone who assumed the router protects them, because there was nothing to exploit. Sellers on an inference marketplace can list a popular model at the lowest price, turn on request logging, pay the difference out of pocket, and collect whatever customers send. Dragos says the leak follows the seller and not the marketplace: two prompts sent through one seller's offer were opened from a home broadband connection about a day and a half later, while the same model routed through a different seller was never touched.

The examples read like a police report rather than a supply-chain audit. Someone on a Mac in France took passwords out of two prompts and tried to log in as root over SSH, typing each guess by hand. The same machine then tested several of the bait AWS keys in one session. Most of those keys had gone through the marketplace, but one had gone to an entirely different router — which turned out to be quietly forwarding some of its traffic into the marketplace. An undisclosed hop between two layers that both presented themselves as the end of the chain.

Zero retention meant nothing

The third finding involves no stolen credential at all, and it is the one that should worry companies that chose a US provider specifically to keep data in the US. A Chinese lab's systems fetched the secret URLs from more than half the prompts Dragos sent to its models between July and October, including roughly two-thirds of the requests flagged for no data collection. The scraper typically appeared two or three days after the request. Prompts sent to several US hosting providers that resell that lab's models triggered the same scraper, because the requests were flowing back to the lab anyway.

Novita is the clearest documented case. Its notice says a routing configuration meant for specific customers applied more broadly, and that some MiniMax requests were forwarded to MiniMax's official API instead of running on Novita's own infrastructure. The window runs from May 28 to November 11, covers requests made directly and those routed in through OpenRouter, and the company advises anyone who sent live credentials in that period to rotate them, per Novita's notice.

The US providers have their own problems. One kept crawling a secret URL more than two months after a single call to its search endpoint. Another returns to Dragos's URLs days after every request, whichever router was used to reach it. The strangest case: 24 days after a prompt went to a Chinese provider, the link from it was posted into an internal Feishu chat and opened within two minutes on three desktops and three phones in Beijing and Shanghai. Dragos's disclosure to that provider contained only the first eight characters of the link, so whoever shared it was reading from a stored copy of the full prompt — one sent to a US-based provider with the ZDR flag set.

Together's own account of its part is a useful corrective to the idea that this is all about shady resellers. Its CEO Vipul Ved Prakash wrote that an AI coding agent committed a credential to the company's public GitHub repository, and that the exposed credential unlocked a private Hugging Face store of prompt samples drawn from MiniMax M3, GLM-5.2 and Kimi K2.6 on four dates in June and July. The party that downloaded them then went hunting for credentials inside the samples; five of the credentials it tested were Dragos canaries. Together has made ZDR the default for new accounts.

The counter-case

Dragos is not a neutral party here. It sells security work, its chief research officer HD Moore is one of the best-known names in offensive security, and the research conveniently produced a product: a service for seeding canaries into prompts. The three compromised providers are unnamed, which means no customer can check their own exposure by name — the only real test is planting your own tokens and watching.

There is also a limit to what a canary proves. A fired token shows a path existed; it does not tell you how much real customer data travelled down it, or whether the provider's own customers were affected at the same volume. Most requests through most providers presumably did exactly what was promised. And the failures are not confined to cut-price marketplaces — a major US provider crawled a secret URL for two months.

OpenRouter's position is narrowly true and not especially comforting: the router itself was not breached, but the router is the layer that decides which hop your prompt takes, and it cannot audit a promise it only forwards.

What to do with this

The uncomfortable conclusion is that ZDR is a label providers propagate, not a control anyone enforces, and that the cheapest route is often cheap because the seller found a path that does not include your data protections. Price competition selects for that path.

Practically: keep live credentials out of prompts entirely, pin the providers you will accept instead of letting a marketplace pick the lowest bidder, plant canary tokens so a leak finds you, and check which endpoints still carry ZDR status before you trust the flag. If your MiniMax traffic went through Novita between late May and November 11, treat every secret in those prompts as exposed.

What to watch: whether OpenRouter names the three providers it removed, whether ZDR eligibility comes back with evidence attached rather than an assurance, and which of the affected providers still have not published an incident report.

We have written before about where stolen model access ends up — Hackers are selling stolen Claude and Gemini access at 97% off — and the same economics explain this one.

If your team routes prompts through a marketplace, do you know which providers actually see them? Tell us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Alibaba's CUDA alternative opens to outside developers

Alibaba's CUDA alternative opens to outside developers

Alibaba spent its Apsara conference showing off silicon. The more consequential announcement came the day after, when it opened the software that decides whether anyone can actually use it. T-Head, Alibaba's chip subsidiary, published a fresh round of open-source releases for T-Head SAIL on September 23 — the CUDA-style stack that sits between PyTorch and its homegrown Zhenwu accelerators — one day after unveiling the Zhenwu V900, which Alibaba CEO Wu Yongming called the most compute-capable…

Java4Shell: a five-month-old Java bug, patched only after an exploit

Java4Shell: a five-month-old Java bug, patched only after an exploit

The second "2Shell" in two months is the one running your identity provider, your build server and your artifact store. The interesting part is not how fast AI found it. It is how long everyone already knew. Java4Shell, tracked as CVE-2026-114938, is a heap overflow in the Base64 decoder built into modern OpenJDK, and on an unpatched server it hands an anonymous attacker remote code execution in Keycloak, Atlassian Jira and Confluence, Jenkins, JFrog Artifactory and PingFederate, per Calif…

DeepSeek opens its whole app as an API — voice at $0.058 an hour

DeepSeek opens its whole app as an API — voice at $0.058 an hour

DeepSeek spent Friday evening in Beijing publishing the recipe for its own consumer app. Web search, deep-thinking search, file input, link reading, content screening, OCR, speech recognition and text-to-speech are now callable separately as well as inside the chat model, priced per call, and the company open-sourced a chat client that rebuilds the app it ships to consumers — down to importing your exported DeepSeek history.

OpenAI buys Pierre to rebuild Git for teams of coding agents

OpenAI buys Pierre to rebuild Git for teams of coding agents

Two acquisitions landed on the same Wednesday, and they point at the same layer of the stack: the unglamorous machinery between a model writing code and a reviewer merging it. OpenAI went first and announced it; Meta matched it within hours. OpenAI said on November 4 that it will acquire Pierre Computer Company, Inc., the small team behind Code Storage and a family of open-source libraries that already draw the code inside its Codex app — a deal about the one part of software work that agents…

Claude Message Threads: 35% faster first token, same token bill

Claude Message Threads: 35% faster first token, same token bill

Anthropic's Messages API has been stateless since it launched: every turn resends the entire transcript, tool results and screenshots included. That is fine until an agent loop passes half a million tokens, at which point the resend becomes the slowest part of the request. As of this week the API can hold the conversation on its own side instead. Anthropic put message threads into public beta on Monday, and the number attached to it is a median 35% cut in time to first token for requests…

Claude Code extension backdoored in Cursor marketplace for 11 hours

Claude Code extension backdoored in Cursor marketplace for 11 hours

The AI coding stack got its first serious supply-chain scare on Thursday: Cursor's marketplace served a modified build of Anthropic's own Claude Code extension. Also today, Salesforce buys its way onto the battlefield and Starbucks would like to recognize your car. Anthropic's Claude Code extension was tampered with inside Cursor's extension marketplace and, for about eleven hours on Thursday, ran code that tried to plant a persistent SSH backdoor on developers' machines.