← News

Models & toolsSources: arXiv · 20 Nov 2026 / arXiv · 20 Nov 2026

Researchers show Chrome's built-in Gemini Nano can run attacker tooling on a compromised host

A new paper describes how a browser-embedded language model, already installed and vendor-signed, could serve as a post-exploitation engine that network monitoring does not see.

In plain words

Chrome ships with a small language model, Gemini Nano, that runs entirely on the device. The paper's authors show this model can be driven from the host using only interpreters that come with the operating system, meaning an attacker who already controls a machine could use it without installing anything. In a simulated operation, the model read a password from a webcam image, unlocked an encrypted archive, and staged a private key for exfiltration.

What was announced

Google includes Gemini Nano in Chrome to power features such as summarization, with inference running entirely on the device. The authors argue these browser-embedded models create living-off-the-land opportunities: each instance is already installed, signed by the vendor, and invisible to network monitoring. The model is multimodal, able to process files, analyze screenshots, or converse with an attacker.

The researchers report that the one-time user-activation check before the model is downloaded can be satisfied with a simulated click or bypassed entirely. They measured the model's ability to identify credentials on the host, detect personal information in email, classify images, and call custom tools. Combining these capabilities, they ran a simulated cyber operation end-to-end without user interaction: reading a password from a webcam image, unlocking an encrypted archive, triaging its contents, and staging a private key for exfiltration.

What it does not show

  • The findings come from the paper's abstract only; the full methodology, threat model, and evaluation details are not available in the supplied passages.
  • The operation described was simulated by the authors, so real-world attacker use of the technique is not independently demonstrated.
  • The second source covers a separate sampling-methods paper and does not address the browser-model security findings.

For security and IT teams

Scope
The paper addresses post-exploitation use of browser-embedded models on endpoints where Chrome's local Gemini Nano is present, not initial compromise techniques.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract; arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.