Inspiration

As users of coding agents, we've singlehandedly felt how annoying it can be when LLMs don't edit the right files, make unnecessary changes, and mess up our projects. That's why we've created AirLock.

What it does

AirLock is a security layer between the prompt and the code that makes security and surveillance easier. Every single path an LLM takes, we categorize and assess the risk. Safe events are automatically passed through, medium risk events require admin approval via iMessage, and extremely dangerous events are blocked before Claude touches the code. And we've reduced the time for surveillance as well: you can now approve/deny tasks, check ongoing tasks and redirect agents, and receive completion messages straight from iMessage.

How we built it

We built a deterministic algorithm in python and built a simple adapter to convert Claude JSON format to a custom dictionary with info most meaningful to us. We created configurable policies in YAML and then we used those polices along with the deterministic algorithm to make decisions on how safe an action might be. Other parts involved properly breaking up the shell of Claude with concatenators like && || etc, building a total deterministic output based on lines/files changes and other factors that went in our model.

Challenges we ran into

Finding a team was initially quite difficult. We initially really liked this idea when we first came up with it, but didn't believe we had the manpower to complete it. We moved around different teams but eventually decided to build our own idea as that was what we were most interested in. Another challenge we ran into was debugging the path of the texts, on some instances it would send out text messages from a different path.

Accomplishments that we're proud of

Building a strong probabilistic algorithm, our python and typescript integration with each other, how we managed to integrate photon after lots of debugging, our collaboration together, the final project and output that we get, the consistency of the output.

What we learned

Learned a lot about what we can do! As both of our first Hackathons, we were surprised with the progress we made and how much more we did than we initially thought. We definitely learned about TypeScript and HTTP connection with Python. We learned how to approach team formation and how to approach building an MVP, prioritising features, etc.

What's next for AirLock

Next for AirLock, we want to continue improving the algorithm by making it team-sensitive and allowing individual teams to manage what they consider risky. We want to package this into an SDK that can be downloaded by others into their own repos and it seamlessly integrates as a safety layer to their phone number for updates. We want to add support for more LLMs and lastly we want to make our risk model more accurate by considering sequences of commands rather than individual commands.