Platanus Hack 26 AI security
yieldOS
A safety net for AI coding agents.
Problem
AI agents can write code fast. Too fast. An agent can install a package, edit a file, and commit, all in seconds. By the time a human reviews it, the damage is done. Teams trust agents to be productive but have no way to say "stop, check this before you commit."
Solution
We built yieldOS: a layer that sits between the agent and your repo. When the agent tries to install a package, edit a file, or run a command, yieldOS checks it against your policies first. Before every commit, an oracle scans for leaked secrets, missing auth, and unsafe changes. If something fails, the commit is blocked with proof of why.
How
Claude Code hooks that intercept agent actions. Git hooks for pre-commit and pre-push. A deterministic oracle that runs checks in 150ms at zero cost. No model calls needed for most checks, just rules. The agent proposes, the oracle decides.
Outcome
Built in 36 hours at Platanus Hack 26. The idea: agents shouldn't be trusted blindly. They should be trusted within boundaries you define.