Why small untested AI-generated diffs on sensitive code paths are the real danger

Short summary

The author identifies a critical failure mode in AI-assisted coding: agents produce minimal, clean diffs that compile and demo well but are untested on sensitive paths (auth, payments, user data). Optimizing for small diffs makes this worse by quietly dropping edge-case handling and failure-path tests. Guardsman, an open-source skill, addresses this by risk-tiering every change, enforcing in-session test execution, and replacing TODO comments with structured triageable watch-log entries.

• •Small untested AI-generated diffs on sensitive paths are more dangerous than large internal scripts
• •Optimizing for minimal diff size increases blast radius by dropping edge-case handling and tests
• •Guardsman risk-tiers changes (trivial→critical), enforces executed tests, and replaces TODOs with structured watch-logs

Generated with AI, which can make mistakes.