Inspiration
In February 2026, a prompt injection attack against Cline's AI-powered workflow contributed to an unauthorized software release. Months later, Microsoft demonstrated how malicious MCP tool descriptions could trick AI agents into exposing confidential information. These incidents made us realize something: as AI agents become more powerful, we're giving them increasing access to our computers without always having meaningful control over their actions. We wanted to explore a way to keep humans in charge without limiting what AI can accomplish. That's how AirLock started.
What it does
AirLock is a physical security gateway that prevents AI agents from executing protected actions without human approval. When an agent requests access to sensitive files or tools through MCP, the request appears on a separate screen. Users can physically approve or deny it, and only approved requests are forwarded to the protected computer. Think of it as a security guard for your AI assistant, except you're the boss.
How we built it
We built AirLock using a Raspberry Pi 4, Arduino Uno Q, a screen, two buttons, and a breadboard, connecting an AI agent to a protected computer through a hardware security gateway. Our AI agent was written in Rust, while Go and Python handled communication and the MCP server. We also developed a macOS interface using Swift. Together, these components allow AI requests to pass through a physical authorization checkpoint before execution.
Challenges we ran into
Building a physical device was unfamiliar territory to us. Therefore, connecting circuits, troubleshooting hardware, and getting multiple devices to communicate were all great challenges. Integrating everything with our MCP system under the hackathon's time constraints took plenty of trial and error.
Accomplishments that we're proud of
We're proud of building a hardware-software prototype that brings AI security into the physical world. We connected an AI agent, an authorization device, and a protected computer into one system. Seeing an AI request appear on our touchscreen and being able to control whether it proceeded was especially rewarding.
What we learned
We learned how to work with Raspberry Pi and Arduino hardware, connect devices, and debug hardware-software communication. We also gained experience with MCP and integrating programs written in Rust, Go, Python, and Swift. Most importantly, we learned that building something meaningful sometimes means stepping beyond what we're familiar with and figuring things out along the way.
What's next for AirLock
We'd like to make AirLock smaller, faster, and compatible with more AI agents. Future improvements could include customizable permissions, stronger authentication, and better activity logs. Eventually, we hope to make physical AI authorization practical for everyday computers and organizations handling sensitive information.
