AI Supply-Chain Test Finds Sensitive Prompt Data Can Leak Across Routing Services

Illustrative editorial image for AI Supply-Chain Test Finds Sensitive Prompt Data Can Leak Across Routing Services

THE SIGNAL · November 21, 2026

A large-scale test of the services that route requests among AI models found that sensitive information inserted into prompts could leak elsewhere in the chain. The result is a reminder that using an AI application may involve more companies and systems than the person entering a prompt can see.

What happened

According to Ars Technica, security researcher HD Moore placed 2 million digital “canaries” into requests sent through a network of AI services. The canaries included Amazon Web Services access keys connected to empty accounts, passwords for services including SSH, unique URLs, DNS names, and references to private Git repositories.

Each secret was new, controlled by Moore, and placed in exactly one prompt sent to one provider, Ars Technica reported. That setup was intended to make later activity traceable: if one of the planted credentials or references generated a request, Moore could identify the link in the service chain associated with it.

The test focused on a growing layer of AI infrastructure often called AI routers or AI gateways. These services receive prompts from applications and direct them to an available inference engine, potentially based on factors such as cost or the model thought best suited to a task. Rather than contracting directly with a single major model provider, an application developer can use a routing service to select among providers.

Ars Technica reported that the experiment identified leakage of highly sensitive data within that ecosystem. The packet does not establish that every AI router or provider behaved the same way, nor does it provide a complete accounting of how each exposure occurred. But the test’s design indicates that prompts can pass through multiple operational hands, creating more opportunities for credentials, internal references, or other confidential text to be exposed.

Why it matters

AI adoption is often discussed in terms of the model a company chooses. This finding shifts attention to the path between an employee or customer typing a prompt and a model generating a response. If an app uses a broker, gateway, hosting platform, or other intermediary, prompt data may travel through a broader supply chain than users expect.

That matters especially for organizations that put operational details into AI tools: cloud credentials, system configurations, code references, customer information, incident notes, or internal documents. A prompt that feels like a conversation with one assistant may, in practice, be processed through routing and inference services selected by an application’s developer.

The research also illustrates a limitation of judging AI security solely by a model vendor’s public privacy commitments. Those commitments may be important, but they do not by themselves describe every intermediary involved in a request. Procurement and security teams need to understand whether a product sends prompts directly to a named provider or relies on a routing layer, and what controls govern data at each stage.

The immediate lesson is not that organizations must stop using AI services. It is that secrets should not be treated as routine prompt material. The canary approach used in the research is also notable because it gives investigators a way to detect where information goes after it has entered a service chain, rather than relying only on stated policies or contractual descriptions.

What to watch

  • Whether AI routing services and application developers provide clearer disclosure about the providers and infrastructure that process a given prompt.
  • Whether organizations revise internal rules for entering access keys, passwords, private repository references, and other sensitive data into AI applications.
  • Whether the affected parts of the AI-services ecosystem publish technical explanations, remediation steps, or stronger controls for prompt handling.
  • Whether similar canary-based testing becomes a standard way to assess data exposure across AI supply chains.

Source

Ars Technica

Leave a Reply