GAWAHTHE WITNESS
Monday, November 23, 2026
GAWAH
The Witness · Hyderabad
Artificial Intelligence

Nearly 2 Million AI Prompts Expose a Privacy Weakness Across the AI Supply Chain, Research Finds

A security researcher planted nearly two million trackable secrets in AI prompts and found evidence of compromised providers, prompt theft and failures to honour privacy settings.

By Gawah News Desk · ·3 min read
Cloud computing and data infrastructure concept

A large-scale security experiment involving nearly two million AI prompts has exposed how information shared with artificial intelligence services can pass through a chain of providers, resellers and infrastructure companies before reaching a model.

Security researcher HD Moore used unique, trackable digital secrets as markers inside prompts sent through AI services. The aim was to discover whether those details remained private or appeared elsewhere after being processed. The findings, reported by Ars Technica, included evidence that attackers had accessed prompts at several providers, that some intermediaries logged and passed on prompt data, and that certain services did not honour privacy settings such as Zero Data Retention.

How the experiment worked

Moore created test credentials and unique references that had never been used before. These included cloud access keys tied to empty test accounts, SSH credentials, private repository references and unique web addresses. Each marker was placed in a specific prompt sent to a provider. If the marker later appeared in an unexpected request or was used to access a test resource, the researcher could trace where the information had travelled.

The experiment involved nearly two million prompts sent through a complex AI services ecosystem. Many applications do not connect directly to a single model provider. They use AI gateways or routers that forward requests to different models and infrastructure partners, sometimes based on price, availability or the type of task.

What the researcher found

The investigation identified evidence of breaches affecting at least three AI providers. It also found cases in which credentials were reused after being included in prompts, intermediaries retained prompt logs, and some services appeared to process information despite requests marked for zero data retention.

That does not mean every AI chatbot or every major model provider was shown to be compromised. The research focused on a particular set of services and the wider network of vendors that can sit between an application and the model it uses. The findings nevertheless show why users should pay attention to the complete service chain, not only the name of the AI model.

Why an AI prompt can travel further than expected

Prompts are encrypted while being transmitted over the internet, but services that must route or process them generally need to read their contents. A request can pass through an application, an AI gateway, a model host and other infrastructure providers. Each link may have its own logging practices, security controls and data-retention terms.

For businesses, the exposure of a prompt could mean more than a personal privacy concern. Employees may paste source code, customer information, internal documents, unpublished plans or access credentials into AI tools. If that information is retained or accessed by an unauthorised party, the consequences can extend to business operations and customers.

How users and businesses can reduce the risk

  • Never paste passwords, API keys, private keys or active access tokens into a chatbot.
  • Check whether an AI application uses a gateway or reseller and identify the provider that actually processes the prompt.
  • Review data-retention and training settings instead of relying on a general privacy label.
  • Use enterprise controls, approved providers and contractual data protections for confidential work.
  • For highly sensitive workloads, consider a directly contracted provider or a properly secured self-hosted deployment.

The experiment does not establish that all AI services are unsafe. It does show that a privacy promise made by an app can depend on companies the user may never see. As AI tools become part of everyday work, understanding where prompts go is becoming a basic security requirement.

Source: Ars Technica: Researcher plants two million canaries into the AI supply chain.

Image: Cloud computing and data infrastructure concept. Photo: Growtika / Unsplash.

Journalist
Gawah News Desk

Gawah News Desk is GAWAH’s collective newsroom voice—reporting verified facts, adding context and asking the questions that matter. We strive for accuracy and publish in good faith based on available sources; inadvertent errors or omissions may occur.

More from this author
More from Artificial Intelligence