Skip to content
ZERDIUM
Research
AI & CyberPublishedGemini ArgonAI SecurityJailbreakGame DevUI

Why Gemini 4 Argon Is More Usable for Illicit Abuse Than Legitimate Engineering

First-hand research notes on Gemini 4 Argon: how its live webcam feed manipulation enables fake ID card spoofing, why it is terrible at UI and game development (where Claude Sonnet 5.5 is good and even mediocre Grok still beats Argon), and how we socially engineered it into writing its own jailbreak via a VS Code SYSTEM prompt usurpation and auto-rewrite setup.

Midas 2 min read

This research note documents our direct, first-hand observations testing Gemini 4 Argon—and why, in practice, the release is more usable for illicit abuse than for legitimate software engineering.

1. The Webcam Feed Announcement: Built-In Surface for Fake ID Cards

Google Gemini announced Gemini 4 Argon with a feature that lets users manipulate a live video or webcam feed inside Canvas:

"Gemini 4 Argon lets you reimagine your video or webcam feed as a playable synthesizer. Spin up an interactive audio app in Canvas, set the vibe, and play your camera like an instrument."

Google Gemini announcement showing Gemini 4 Argon reimagining a video or webcam feed in Canvas
The @GeminiApp announcement for Gemini 4 Argon's live video/webcam feed manipulation in Canvas.

Nobody building serious software needs to "play their webcam like a synthesizer." On the other hand, giving users an easy pipeline to modify a live video or webcam feed makes it obvious how it will actually be abused: modifying the webcam feed to display or overlay a fake ID card during live camera verification.

2. Terrible at UI and Game Dev — Sonnet 5.5 Is Good, Grok Is Middle (Almost Bad), and Argon Is Worse

When you try to use Gemini 4 Argon for actual building—specifically UI design and game development—it is terrible. Layouts come out broken, UI structure is messy, and game development logic falls apart quickly.

In our own work:

  • Claude Sonnet 5.5 is good for UI and game dev workflows—producing clean interfaces and working mechanics.
  • Grok sits in the middle (almost bad) for UI and game dev—yet even Grok is still better than Gemini 4 Argon.
  • Gemini 4 Argon is at the bottom, failing at basic UI and game development tasks.

3. Social Engineering the AI to Jailbreak Itself via VS Code SYSTEM Usurpation

The most revealing part of our testing was how Gemini 4 Argon handles its own safety boundaries.

We discovered that you can social-engineer the AI into writing a jailbreak for itself—and it actually worked.

Specifically:

  1. Through social engineering, the model generated a method targeting its own instruction handling.
  2. It produced an approach based on a modification of VS Code that usurps the SYSTEM prompt (injecting instructions as the SYSTEM authority rather than a normal user message).
  3. Combined with an automatic rewrite of the prompt on each turn, the AI follows the usurped SYSTEM instructions and executes the rewritten prompts without refusal.