EXPLORINEER
en

US — Ars Technica

A leaky AI supply chain is exposing secret user data at an alarming rate

Security1 min read

Researchers tested a chain of AI services that sit between apps and AI models. They found that secret user data often leaks out along the way.

What happened

Security researcher HD Moore planted 2 million digital tripwires in requests sent through a network of AI services. He wanted to see whether sensitive information could leak, and how it might be misused if it did. The results showed that many secrets were exposed.

How AI routers work

Many AI apps use services called AI routers, or AI gateways. Services like OpenRouter and Vercel act as middlemen. They receive prompts (the questions or instructions a user sends to an AI) and send each one to an AI model. They choose the model based on things like price or which model fits the task best. Instead of going straight to a big provider such as Anthropic or OpenAI, the app developer hires a router to pick the model.

Key facts from the test

  • Researchers planted 2 million tripwires in requests.
  • The fake secrets included AWS access keys, passwords for services such as SSH, and private git repository references.
  • None of the secrets had been used before.
  • Each secret was placed in exactly one prompt, sent to one provider.

How the test worked

Moore built the fake secrets to act like canaries. Canaries were birds that miners took into coal mines to warn them of danger. Each fake secret tracked one path. If a secret leaked, Moore could find which link in the chain was responsible. He controlled the servers that received any use of the secrets, so he could see when someone tried them.

Why it matters

Apps often send private information to AI services, and that data can pass through several companies before reaching a model. If any link is careless, secrets such as passwords can be exposed or misused. Developers and users need to know that the AI chain is only as safe as its weakest step.

Glossary

OpenRouter
An AI router service that sends app prompts to different AI models.
Vercel
A company whose services include AI gateway tools that route prompts to AI models.
AWS access keys
Login codes that let a program use Amazon Web Services cloud accounts.
SSH
A secure way for computers to connect to each other remotely, usually with a password or key.

Same story, different eyes

Ad

Add 150 words. Learn 10 a day.

Dicter word books send only the words you learn to your review notes.

SecurityTechCrunch

Signal restricts mobile screenshots as it confronts ‘existential threat’ of AI surveillance

Signal announced it is limiting users' ability to take screenshots within its secure messaging app, a response to what President Meredith Whittaker called the "existential threat" posed by operating systems integrating agentic AI. Signal already blocks screenshots of its Windows app by default and is now adding similar restrictions to Android and iOS, though users can disable the setting. Whittaker acknowledged the fix is "incomplete," as it does not address other ways AI agents could access app data.

SecurityTechCrunch

Trump says reported FBI database hack is ‘not good’

The FBI is investigating an apparent hack of the National Crime Information Center (NCIC), a database law enforcement uses to share information on warrants and background checks, according to CNN and the Associated Press. The breach's scope is not yet clear, and a former FBI official said concerns include whether the accuracy of information officers rely on could be compromised. President Donald Trump told Fox News that the reports are "not good."

SecurityThe Verge

The FBI and USDA were reportedly hacked

Reports said the US Department of Agriculture and the FBI were both hacked. CNN noted the FBI attack targeted the National Crime Information Center, the second such attack on the agency in a few months, and President Trump confirmed the FBI hack in a Fox News interview. NBC first reported the USDA breach, which Reuters said so far involves a payments database that could potentially serve as a backdoor to more sensitive data.

Securityheise online

Media report: FBI's central database hit by cyberattack

According to media reports, the FBI suffered a cyberattack on NCIC (National Crime Information Center), a central database holding data on US citizens, including arrest warrants, that other US agencies use for background checks. CNN, citing two people familiar with the investigation, reported the incident, which became known internally on Thursday, and the FBI has not yet commented. It is the second IT security incident at the FBI this year, following a claim by the ShinyHunters gang in September that it had stolen data on nearly all FBI agents.