Chrome's Gemini Nano abused as a post-exploitation engine
Original title: Living off the Local Model: Abusing Browser-Embedded Language Models for Post-Exploitation
Authors used only OS interpreters to have the on-device model read a webcam password and stage a private key.
A paper shows Google's on-device Gemini Nano in Chrome can be driven after compromise using interpreters that already ship with the operating system. The authors say a simulated click satisfies the one-time download activation check, or the check can be bypassed. They report the model can find host credentials, detect email PII, classify images, and call custom tools, then demonstrate a simulated operation that unlocks an archive and stages a private key without user interaction.
Unconfirmed: The end-to-end operation is an author simulation, not a confirmed real intrusion.
- Inference for features such as summarization runs entirely on the device.
- The model is multimodal and can process files, screenshots, and attacker conversation.
- The demonstrated operation triaged an unlocked archive and staged a private key.
- No user interaction was required for the end-to-end simulation.
Impact: On-device model finds credentials and stages keys
| Product | Affected |
|---|---|
| Google Chrome Gemini Nano | — |
- Vendors
- Products
- Chrome
- AI models
- Gemini Nano
From the original article
1 min read at arxiv.orgBrowsers now ship local language models to ordinary endpoints. Google includes Gemini Nano in Chrome to enable productivity features such as summarization, with inference running entirely on the device.
By William Fleshman · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.