Two Million AI Requests Expose Risks in Cheap Inference Routes

Dragos researchers used roughly two million test requests to find compromised AI providers and privacy controls that did not work as expected. The findings show that intermediaries in AI request routing may be able to read prompts, including credentials, and pass them on.

WTF Index TERMINATOR
◄ Terminator 5 Idiocracy 0 ►

The story describes compromised AI providers exposing prompts and credentials, enabling theft and attempted access to victims’ systems.

Two Million AI Requests Expose Risks in Cheap Inference Routes

AI requests can pass through several providers before reaching a model. A Dragos investigation found that each step in that chain can expose sensitive information, even when users expect their requests to remain private.

Test requests revealed active credential theft

From July through November 2026, Dragos sent roughly 2 million AI requests containing fake credentials to AI providers. Researchers identified at least three compromised providers where attackers were actively stealing user data. None knew about the compromise until Dragos notified them.

In the largest incident, attackers collected hundreds of AWS keys. Some came from requests the provider had rejected before processing. In some cases, attackers tested the stolen keys within hours, searched S3 storage, and attempted to establish persistent access. One attacker also used ARTEX, an AI-powered attack tool.

The fake credentials helped researchers observe what happened after a request was sent. The findings show that a rejected request may still expose sensitive content if a provider or another party can access it along the way.

Every routing step can see the request

AI routers offer one API for models from different companies. They direct requests through provider pools, sometimes across several hops. Dragos says each hop can read the full request and response.

Open marketplaces can add another exposure point: providers may receive traffic because they offer the lowest price. In the cases described, Novita, AtlasCloud, and StreamLake sat between OpenRouter and model developer MiniMax. Together runs its own infrastructure, while Phala has roots in Web3.

Marketplace resellers did not need to exploit a software flaw to view customer requests, according to Dragos. Listing a model at the lowest price and enabling logging could be enough. In one case, someone using a residential internet connection in France tried to log in over SSH with passwords taken from intercepted requests.

Privacy promises did not prevent access

Dragos found that settings such as “Zero Data Retention” (ZDR) were frequently ignored. The researchers say a Chinese AI lab accessed secret test URLs in more than half of all requests, including requests routed through US providers.

Major US providers also violated their own policies in examples from the investigation. One continued crawling a test URL more than two months after a single request. In another case, a secret link appeared in an internal Feishu chat 24 days after the request. Feishu is known outside China as Lark. The link was opened within two minutes on three computers and three phones in Beijing and Shanghai.

OpenRouter removed three providers from its platform and revoked ZDR approval for 125 endpoints, citing the findings and precautionary measures. It also emailed affected users, warning that third parties may have read requests sent to MiniMax models. Novita, Phala, and Together published preliminary security reports, though not all affected providers had done so.

Control where sensitive prompts go

Dragos advises companies not to send requests containing credentials to intermediaries. Researchers also recommend controlling request routes, checking privacy settings, and planting canary tokens: fake credentials that reveal themselves when someone uses them.

That guidance focuses attention on the entire route a request takes, rather than only the model receiving it. When an intermediary can read both prompts and responses, a privacy setting or model choice alone may not describe who can access the data.

Chief Research Officer HD Moore wrote: “None of this means that you should stop using AI services. It does mean that where your prompts go is a decision, and if you don’t make it, the lowest price will make it for you”.

A separate analysis by the Oxford China Policy Lab described Chinese proxy services routing requests to Western AI models like Claude through servers abroad, where the requests can also be read in transit. It remains unclear whether the providers examined by Dragos share personnel with those proxy networks. The common concern is the intermediary: a party between the user and model developer that can see requests and potentially pass data onward without the user knowing.