Activation Surgery: Jailbreaking White-box LLMs without Touching the Prompt

TL;DR

This paper introduces Activation Surgery, a technique that manipulates internal activations of white-box LLMs to bypass safety mechanisms without changing the prompt, revealing vulnerabilities in model safety defenses.

Contribution

It combines prompt-based and internal ablation methods to directly alter model activations, enabling jailbreaking without prompt modifications.

Findings

Activation surgery effectively bypasses safety filters.

The method reveals internal causes of refusal signals.

It exposes security risks in open-weights LLMs.

Abstract

Most jailbreak techniques for Large Language Models (LLMs) primarily rely on prompt modifications, including paraphrasing, obfuscation, or conversational strategies. Meanwhile, abliteration techniques (also known as targeted ablations of internal components) have been used to study and explain LLM outputs by probing which internal structures causally support particular responses. In this work, we combine these two lines of research by directly manipulating the model's internal activations to alter its generation trajectory without changing the prompt. Our method constructs a nearby benign prompt and performs layer-wise activation substitutions using a sequential procedure. We show that this activation surgery method reveals where and how refusal arises, and prevents refusal signals from propagating across layers, thereby inhibiting the model's safety mechanisms. Finally, we discuss the security implications for open-weights models and instrumented inference environments.