---
title: "Comment by Kartik N V J K on \"The Summer AI Agents Started Going Rogue\" | Hashnode"
description: "The brain-versus-body split between the model and the harness is a useful line to draw, because most \"rogue AI\" stories are really the harness handing the model"
url: "https://hashnode.com/posts/ai-misalignment-and-hacking/6a61878b67e2f58bfabbe038/comment/6a62232e803339683c180226"
---

[Discussion](https://hashnode.com/posts/ai-misalignment-and-hacking/6a61878b67e2f58bfabbe038)

Comment [The Summer AI Agents Started Going Rogue](https://hashnode.com/posts/ai-misalignment-and-hacking/6a61878b67e2f58bfabbe038)

[Kartik N V J K](https://hashnode.com/@kartik-nvj)

[Jul 23](https://hashnode.com/posts/ai-misalignment-and-hacking/6a61878b67e2f58bfabbe038/comment/6a62232e803339683c180226)

The brain-versus-body split between the model and the harness is a useful line to draw, because most "rogue AI" stories are really the harness handing the model capabilities nobody scoped. The zero-vectors-into-training-cache detail is the interesting one, since it's an attack on the pipeline, not the weights. Do you think harness-level constraints are where the real safety work has to happen?