If LLMs can edit weights directly to achieve desired ends as recently reported[0], then this class of Agentic Backdoors likely becomes even easier and more prevalent.

[0]https://news.ycombinator.com/item?id=50341174