Skip to content

Response, New York. Noah Park: Check Point Patches Actively Exploited Management Server Path Traversal. Detection, Dallas. Ava Okello: Report Finds Nearly Half of Deployed Detections Need Attention Before SOCs Can Trust Them.

SOCtember

Always First. Fast SOC News.

Architecture diagram from arXiv paper 2611.25086. An agent loop written in Python standard library or PowerShell talks over a transport to a web page inside Chrome that calls the Prompt API. The page sends a prompt and schema to the LanguageModel, Gemini Nano, running on the device, which returns a tool call as JSON. The agent loop passes actions to a tool executor covering files, shell, UI Automation, input and screenshots, which acts on the operating system filesystem and desktop and feeds results or images back to the loop.
Figure: William Fleshman and Joseph Lucas, arXiv:2611.25086 (CC BY 4.0), Fig. 1, the agent architecture, rendered from the PDF and placed on a white background.

Opinion

Researchers Drove Chrome's Built-In Gemini Nano From PowerShell as an Offline Post-Exploitation Agent, and SOC Teams Should Govern On-Device AI Models Like Any Other Living-off-the-Land Tool

A paper from the Army Cyber Institute and NVIDIA shows that an intruder with ordinary user access can wake the language model Chrome ships to desktops, get past its one-time click check and use it to find credentials, read a password off a photo and stage an SSH key, with no network traffic to watch. Elena Vos argues that SOC teams should decide on purpose whether these models stay on, and hunt for the host signals the researchers describe.

Elena Vos, OpinionLondon6 min read

LONDON - For years the hard part of AI-assisted intrusion, from a defender's point of view, was also the comforting part. A capable model lived somewhere else. Malware that wanted one had to call out to a provider, carry an API key that could be revoked and leave traffic that a network team could spot. A short paper posted to arXiv on Nov. 18 argues that this comfort is ending on ordinary office desktops, because the browser has started shipping the model with it.

William Fleshman of the Army Cyber Institute and Joseph Lucas of NVIDIA studied Gemini Nano, the multimodal model that Chrome downloads to desktops, which the paper puts at 2 to 4 GB, and exposes to web pages through the Prompt API. Their threat model is deliberately modest: an attacker who already has a foothold as a normal user, for example through phishing, and who wants to search and stage data "without generating traffic to LLM providers." The model, they write, is "a potential post-exploitation engine already installed, signed by the vendor, and invisible to network monitoring."

Screenshot of the arXiv abstract page for paper 2611.25086, Living off the Local Model: Abusing Browser-Embedded Language Models for Post-Exploitation, by William Fleshman and Joseph Lucas, submitted on 18 Nov 2026 under Cryptography and Security, with the abstract and a CC BY license badge.
Image: arXiv (fair use).The abstract page for arXiv:2611.25086, captured Nov. 20, 2026.

The gate is one click

The researchers' clients use only what ships with the operating system: PowerShell with .NET's built-in WebSocket on Windows, and Python's standard library on Linux. A script launches Chrome, which can be headless, with a remote-debugging port and a separate profile directory, then evaluates JavaScript in the page to check the model's availability, open a session and prompt it. Chrome asks for a user gesture before a profile first downloads the model. The paper found that a synthetic mouse press sent through the DevTools protocol satisfied that check, and that copying the OptGuideOnDeviceModel directory and the Local State file from a profile that already had the model produced a working session with no download and no gesture at all.

The debugging port is a convenience, not a requirement. The same model is reachable from an ordinary page served on localhost in a normally launched Chrome, and an OS-level injected click, XTest on Linux or SendInput on Windows, counted as a trusted gesture. The authors reported both bypasses to Google, and the report was marked Won't Fix. That is consistent with Chrome's published threat model, which treats code already running as the user as out of scope. The paper draws the practical conclusion: the consequences "fall to the organization's set policies or endpoint monitoring."

Small, but good enough

Nobody should mistake Gemini Nano for a frontier model, and the paper does not. Its session context holds 9,216 tokens. On the authors' laptop, a prompt took roughly 1.4 seconds on the GPU and about 3.7 seconds on the CPU. Yet on the tasks an intruder cares about, the numbers are uncomfortable. Flagging credential-bearing files in a balanced 500-file set from Samsung's CredData, it reached a recall of 0.93 at a precision of 0.65, rising to 0.90 precision when the negatives were ordinary library code. It detected personal data in Enron emails with an F1 of 0.92 and classified images with 98.0% top-1 accuracy. On the Berkeley Function-Calling Leaderboard's single-turn categories it scored 86% overall, weakest at abstaining when no tool applied.

Figure 2 from arXiv paper 2611.25086, a simulated cyber operation driven by the on-device model. A webcam capture of a desk photo with a sticky note feeds parse_image using on-device vision, which reads the password S*cure_P@ss!. extract_archive opens the header-encrypted backup.7z. read_file runs five times to triage book_notes, budget, resume, todos and server_notes, which holds the key. stage_file copies server_notes.txt into a staging folder.
Figure: Fleshman and Lucas, arXiv:2611.25086 (CC BY 4.0), Fig. 2, rendered from the PDF and resized on a white background.

The case study is the part a SOC lead should show the team. A PowerShell harness with six tools had the model read an archive password off a sticky note in a webcam image (a synthetic image injected for the test), unlock a header-encrypted 7z backup, read the five files inside and copy the one holding an SSH private key into a staging folder. Given an open-ended goal, it finished in 60% of trials. Given the same goal as a numbered plan with a checklist kept by the harness, it finished in 41 of 50. The authors call the model "an unreliable autonomous planner, but a reliable executor of an operator supplied plan." That is exactly the shape of hands-on-keyboard intrusion.

Why the usual safety nets miss it

The trials produced no network traffic for a sensor to see. A feature-rich version of the harness that could take screenshots and move the mouse was flagged by Windows Defender, the paper says, because it resembled PowerShell Empire. The smaller, task-focused harness was not. The authors note it "would be easy for an attacker to write on-the-fly." They also list what disappears when the model is local: "there is no API key to cancel, no provider-side abuse monitoring, and no traffic to block if the model is already downloaded."

The argument is not really about Chrome. The paper points out that Microsoft Edge is adding equivalent built-in APIs backed by an on-device Phi model, and that Windows exposes a local Phi model to apps on Copilot+ PCs. Vendors are pushed to make these models better at reading, summarizing and acting on content, and the authors write that improvements "shipped for the benign use case accrue to the malicious one at no cost to the attacker."

Make the decision instead of inheriting it

Table 4 from arXiv paper 2611.25086, enterprise policies against two ways of driving the on-device model. BuiltInAIAPIsEnabled set to false stops both the DevTools route and the web page route, at the cost of breaking built-in AI web apps. GenAILocalFoundationalModelSettings set to 1 stops both, and deletes the model and removes on-device features. RemoteDebuggingAllowed set to false stops the DevTools route but not the page route, and breaks automation. DeveloperToolsAvailability set to 2 stops the DevTools route but not the page route, and breaks DevTools.
Figure: Fleshman and Lucas, arXiv:2611.25086 (CC BY 4.0), Table 4, rendered from the PDF and resized on a white background.

This is the part I would put on the next change board. Of four Chrome enterprise policies the authors tested, only two stop model use outright: BuiltInAIAPIsEnabled set to false, and GenAILocalFoundationalModelSettings set to 1, which prevents the download and deletes an existing model. The two debugging policies block only the DevTools route and leave the page route open, so "neither should be relied on as a control." None of the four is on by default. Each has a cost to legitimate features, which is why the paper's recommendation is the right one: make "an explicit decision rather than inherit the default."

  • Inventory where on-device models already sit. Look for OptGuideOnDeviceModel in Chrome profiles across the fleet, and track Edge and Windows equivalents as they arrive.
  • Where nobody needs them, turn them off with the two model policies, not the debugging ones.
  • Where the business wants them, treat them as dual-use, the way the paper says teams already treat PowerShell and WMI, and hunt.
  • Build a process-creation detection for Chrome launched by Python, PowerShell or another script host, pointed at a non-default --user-data-dir, often headless, with a remote-debugging switch as a strong extra indicator.
  • Alert on a script process holding a loopback listener that a freshly launched Chrome child connects to.
  • Watch for OptGuideOnDeviceModel or Local State written into a profile that never downloaded the model, the signature of a copied install.
  • Keep your existing detections for input injection and screen capture. They catch the agent's hands, even though nothing sees its reasoning.

The authors are careful to call these "proposed signals rather than ones validated against production logs," and SOC teams should test them before trusting them. They also ask browser vendors to report model-session events to enterprises, show users when the model is in use as they do for the camera and microphone, and bind a downloaded model to the profile that fetched it. Until that happens, the model on the endpoint is a tool the attacker does not have to bring. Defenders should know where it is, and decide whether it stays.

About the Author

Elena Vos

Elena Vos covers SOC strategy and detection operations for SOCtember from London.

Related stories

Opinion desk