Vulnerability / research
Researchers use Chrome’s local model in a simulated post-exploitation chain
Reported
Published here
What happened
Researchers with pre-existing ordinary-user access to one test laptop drove Chrome’s Gemini Nano model from local scripts. Their controlled scenario read a password from a synthetic webcam image, opened an archive, identified a staged private key and copied it to a staging directory.
Parties and reporting
Where this happened
Research, tests and decoys. Lab studies, demonstrations, sandboxed evaluations, honeypots and disclosed flaws without a documented real-world victim. There is no victim location to map. Explore the world map.
AI tools and systems: Chrome 156 on Ubuntu / WSL2, Chrome 157 on Windows 11, Gemini Nano Prompt API, PowerShell and Python research harnesses
Outcome
An explicit plan and checklist completed 41 of 50 trials; an open-ended goal completed 60% of 20 trials. Researchers say Google marked the reported model-download gesture behavior Won’t Fix because code already running as the user is outside the browser’s security boundary. No production intrusion or external data theft was demonstrated.
Operating lesson
Treat local models as capabilities available to compromised user processes, make an explicit enterprise policy decision about model access, and monitor the surrounding process and file activity.
Evidence and limits
This is controlled post-exploitation research, not a remote initial-access exploit or a new asserted product vulnerability. The attacker already has the user’s execution privileges; Chrome’s process sandbox remained enabled. The webcam image and target files were staged, and the result was local staging rather than observed exfiltration. A richer harness triggered Windows Defender while a narrower harness did not, so universal evasion is not established. Google’s disposition is reported by the authors; the November 10 issue was restricted and was not independently reviewed. November 18 is the earliest verified public paper.
Sources
- Full original study: threat model, single-laptop tests, outcomes and policy limits
Primary source ·
- Original arXiv submission history
Primary source ·