Wrapped up OpenAI Build Week with another round on TraceTree, a runtime behavioral security scanner I’ve been building outside of any competition. It uses radare2, YARA, and an ML pipeline to catch malicious files and surface the verdict in a dashboard. I bring it to hackathons that allow existing projects because it’s become my favorite way to learn a new stack without throwing away the work after.
This round the stack was Codex with GPT-5.6, and the fix that mattered most was a quiet one. The backend was never emitting a “malicious” verdict in certain cases, so genuinely dangerous scans were showing up as a yellow caution instead of red danger. Codex helped me trace that bug across the orchestrator, the server, and the frontend before I nailed down the fix.
The real lesson every hackathon reteaches me is what to hand an agent and what needs my own judgment. A wide open scan burns credits with barely any signal. A narrow, scoped ask gets you something you can actually verify.
TraceTree keeps growing one hackathon at a time, and I keep learning something different about working with these tools each round.