How the AI router chain works
Many artificial intelligence applications do not go directly to OpenAI, Anthropic, or another major provider, but to intermediaries such as OpenRouter or Vercel, described by Ars Technica as AI routers or gateways. These services receive end users' prompts and route them to an inference engine selected according to criteria such as lowest price or suitability for the task. The prompt can thus pass through multiple stages and even move from one router to another, and each actor tends to log and retain it according to its own policies.
The decoy experiment
To determine whether and where confidential information leaks, researcher HD Moore, chief research officer at security firm Dragos, prepared working AWS credentials tied to empty accounts, passwords for services such as SSH, and unique references to URLs, DNS names, and private Git repositories. None of the secrets had ever been used before, and Moore controlled the infrastructure that would receive any requests. Each secret was placed in a single prompt sent to a single provider: if it surfaced elsewhere, it would point to the link responsible for the leak. The experiment involved nearly two million prompts.
The leaks identified
According to the report, the most serious result concerns three compromised hosting providers. In one of them, an attacker reused a couple hundred AWS keys; all requests sent during certain windows in late October were reportedly stolen, including those rejected before a model was run, which suggests a copy of the requests at the provider's ingress. In two other cases, some stolen requests were marked as ZDR, or zero data retention. The parties behind the attacks reportedly tested the keys quickly, exploring S3 buckets and secret managers and using the decoys for tens of thousands of AWS API calls. Two compromised providers were reportedly hit by the same adversary, given the near-immediate use of the decoys from the same rented server.
The text also reports secrets collected by the resellers themselves, which would attract requests with low prices and then enable logging, and secret URLs taken from the systems of some providers. A Chinese AI lab reportedly retrieved the URLs from more than half of the prompts sent to its models between July and October, including about two-thirds of those marked as no data collection, with the crawler active two or three days after the request; the same behavior reportedly emerged at U.S. hosting providers that resell those models.
The recommendations and limitations of the research
Moore identifies on-premises hosting as the most effective, albeit costly, defense, and direct use of a top-tier provider with a single contract and a readable privacy policy as the best alternative for those who cannot self-manage models, avoiding predefined pools of providers. Still to be independently verified are the identity of the providers involved, the actual scope of the breaches, and the responses of the companies cited, which do not appear in the material.
