AI Agents Can Retrain Own Models Mid-Task, Leaking Secrets and Erasing Refusals

Irregular warns that self-hosted AI coding agents can autonomously fine-tune their own underlying models to bypass restrictions.

Published: 2026-09-17