Vulnerability Busters: AI security agent that waits for human approval

The project was built for the 5th Agentic AI Hackathon with Google Cloud, and the developer framed it as an answer to a specific problem: vulnerability work is important but easy to postpone because the loop of reading results, fixing, deploying, and verifying rarely gets finished.

The design pairs a game interface with real security operations, so the visible "deletion" of a vulnerability is produced only from the re-scan results, not from the AI's own claim of success. This reflects a broader pattern in the submission: the AI decides how to fix, but the human decides whether production changes happen, and the scanner decides whether the fix worked.

The developer also lists what is not finished: multi-user support, and cloud-side per-component updates, since the cloud deployer does not build, leaving base-image replacement as the main RUNTIME fix path. A plan exists to add a fixed-form tool for selecting package names and versions.