Research Notes

Long-form notes about experiments, failed hypotheses, evaluation design, and lessons that transfer beyond a single project.

AI agent security

AI Agent Security After 29 Attempts

A source-first postmortem on 28 scored Kaggle submissions, one system error, replay-order confounding, public-leaderboard overfitting, and the private-guardrail reset.

Published September 2, 2026 · 11 minute read