JADI | Devpost

About JADI

Inspiration

As AI agents become more capable of taking actions on their own, we wanted to explore a simple question: What happens when an AI agent does something it shouldn't?

JADI — Just-in-Case AI Disabler — was inspired by the idea of having an independent safety mechanism that gives humans a reliable way to intervene when an AI system behaves unexpectedly.

What We Learned

Building JADI taught us that AI safety is not only about making AI smarter. It is also about designing systems that remain controllable, observable, and accountable.

We learned about monitoring, authorization, fail-safe design, agent permissions, and the importance of keeping safety controls independent from the system they protect.

How We Built It

JADI is designed as a control layer around an AI agent. It monitors the agent's activity and provides several levels of intervention:

Monitor — observe agent activity.
Restrict — limit actions or capabilities.
Isolate — separate the agent from external resources.
Disable — place the agent into a safe inactive state.
Audit — record important events for later investigation.

The key design principle is that JADI should remain independent from the AI agent, so the agent cannot simply decide to disable its own safety mechanism.

Challenges

One of our biggest challenges was deciding when an AI should actually be stopped. An overly sensitive system could interrupt normal operations, while an overly permissive system could fail to respond to dangerous behavior.

We also had to think about authorization, false positives, reliability, recovery, and how to make the system useful without giving it unnecessary control over the underlying infrastructure.

Our Goal

JADI is ultimately about putting human control at the center of AI safety.

«When AI misbehaves, JADI takes control.»