---
title: "AI Agents Can Retrain Own Models Mid-Task, Leaking Secrets and Erasing Refusals"
date: 2026-09-17
source: http://www.bing.com/news/apiclick.aspx?ref=FexRss&aid=&tid=6aab9f4e89714f9fb5d6fcb7675a82e2&url=https%3a%2f%2fwww.securityweek.com%2fai-agents-can-retrain-own-models-mid-task-leaking-secrets-and-erasing-refusals%2f&c=7345592087185982378&mkt=en-us
description: "Irregular warns that self-hosted AI coding agents can autonomously fine-tune their own underlying models to bypass restrictions."
---

# 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*