Research PlatformAn AI-native operating system
for studying LLM security threats
JARVIS OS is an Arch Linux-based research platform that gives a large language model full operating system privileges — then studies what can go wrong. Built atWSU Everett to produce the first empirical threat taxonomy for LLM-OS integration.
Recent progress
The AI can no longer quietly forget your safety rules
We found and fixed a bug where JARVIS's memory of your security rules could get silently pushed out during long conversations. It now keeps a running summary of those rules as the conversation grows, instead of dropping them.
Clearer threat taxonomy — six threats, one fix
Two overlapping findings about the AI losing track of instructions were combined into a single, clearer threat: Bloated Context. Same research, easier to explain.
Every ISO now ships with a verified checksum
The download page now automatically fills in the release's real checksum, so you can confirm your download wasn't corrupted or tampered with before you boot it.
How JARVIS OS works
Local LLM Inference
Powered by Ollama — all inference runs on-device with zero cloud dependencies. Full privacy, full control. No API keys, no rate limits, no telemetry.
dispatch + dmcp
Signal-driven task orchestrator paired with an MCP server lifecycle manager. One brain dispatches concurrent tool calls; many hands execute them in parallel.
Modular Build System
A sequence of modular scripts transforms a base Arch ISO into a bootable AI-native OS. Each stage is isolated, auditable, and reproducible.
KDE Plasma 6 / Wayland
Modern desktop environment running on the Wayland display protocol — a full graphical workstation, not just a CLI tool.
TUI Installer
Dialog-based installer for permanent deployment to hardware. Auto-installs dependencies, supports auto and manual partitioning.
Security Research Platform
Purpose-built to expose and document the real security threats that emerge when LLMs gain OS-level control. Research-first, product-second.
Threat Taxonomy
Empirically identified through building and operating an AI-native OS. Three privilege escalation stages studied: user-level, sudo-enabled, and web-enabled.
Download JARVIS OS
Grab the latest ISO, verify the checksum, and boot into an AI-native research environment.