Cover Image

The Rogue Scalpel: Activation Steering Compromises LLM Safety

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Korznikov, Anton, Galichin, Andrey, Dontsov, Alexey, Rogov, Oleg Y., Oseledets, Ivan, Tutubalina, Elena
Format:	Preprint
Published:	2025
Subjects:	Machine Learning Artificial Intelligence
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

Similar Items

OrtSAE: Orthogonal Sparse Autoencoders Uncover Atomic Features
by: Korznikov, Anton, et al.
Published: (2025)

Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?
by: Korznikov, Anton, et al.
Published: (2026)

GLiRA: Black-Box Membership Inference Attack via Knowledge Distillation
by: Galichin, Andrey V., et al.
Published: (2024)

I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders
by: Galichin, Andrey, et al.
Published: (2025)

Spread them Apart: Towards Robust Watermarking of Generated Content
by: Pautov, Mikhail, et al.
Published: (2025)

GigaEvo: An Open Source Optimization Framework Powered By LLMs And Evolution Algorithms
by: Khrulkov, Valentin, et al.
Published: (2025)

Steered LLM Activations are Non-Surjective
by: Mishra, Aayush, et al.
Published: (2026)

BarrierSteer: LLM Safety via Learning Barrier Steering
by: Tran, Thanh Q., et al.
Published: (2026)

Steer Like the LLM: Activation Steering that Mimics Prompting
by: Heyman, Geert, et al.
Published: (2026)

Graph-Regularized Sparse Autoencoders for LLM Safety Steering
by: Yeon, Jehyeok, et al.
Published: (2025)

Steering LLM Reasoning Through Bias-Only Adaptation
by: Sinii, Viacheslav, et al.
Published: (2025)

Motivating Next-Gen Accelerators with Flexible (N:M) Activation Sparsity via Benchmarking Lightweight Post-Training Sparsification Approaches
by: Alanova, Shirin, et al.
Published: (2025)

Your Transformer is Secretly Linear
by: Razzhigaev, Anton, et al.
Published: (2024)

On the Spatial Structure of Mixture-of-Experts in Transformers
by: Bershatsky, Daniel, et al.
Published: (2025)

Exploring the Hidden Capacity of LLMs for One-Step Text Generation
by: Mezentsev, Gleb, et al.
Published: (2025)

Binding threshold units with artificial oscillatory neurons
by: Fanaskov, Vladimir, et al.
Published: (2025)

Quasi-Random Physics-informed Neural Networks
by: Yu, Tianchi, et al.
Published: (2025)

Bayesian Inverse Problems Meet Flow Matching: Efficient and Flexible Inference via Transformers
by: Sherki, Daniil, et al.
Published: (2025)

Scalpel vs. Hammer: GRPO Amplifies Existing Capabilities, SFT Replaces Them
by: Rajani, Neel, et al.
Published: (2025)

Dynamically Scaled Activation Steering
by: Ferrando, Alex, et al.
Published: (2025)

The Ky Fan Norms and Beyond: Dual Norms and Combinations for Matrix Optimization
by: Kravatskiy, Alexey, et al.
Published: (2025)

CLEAR: Character Unlearning in Textual and Visual Modalities
by: Dontsov, Alexey, et al.
Published: (2024)

The Structural Scalpel: Automated Contiguous Layer Pruning for Large Language Models
by: Lu, Yao, et al.
Published: (2025)

Latent Knowledge Scalpel: Precise and Massive Knowledge Editing for Large Language Models
by: Liu, Xin, et al.
Published: (2025)

Activation Steering for Chain-of-Thought Compression
by: Azizi, Seyedarmin, et al.
Published: (2025)

Minimizing Collateral Damage in Activation Steering
by: Nguyen, Tam, et al.
Published: (2026)

Risk Awareness Injection: Calibrating Vision-Language Models for Safety without Compromising Utility
by: Wang, Mengxuan, et al.
Published: (2026)

HyperSteer: Activation Steering at Scale with Hypernetworks
by: Sun, Jiuding, et al.
Published: (2025)

The Shape of Learning: Anisotropy and Intrinsic Dimensions in Transformer-Based Models
by: Razzhigaev, Anton, et al.
Published: (2023)

Global Evolutionary Steering: Refining Activation Steering Control via Cross-Layer Consistency
by: Jiang, Xinyan, et al.
Published: (2026)

Certification of Speaker Recognition Models to Additive Perturbations
by: Korzh, Dmitrii, et al.
Published: (2024)

Probabilistically Robust Watermarking of Neural Networks
by: Pautov, Mikhail, et al.
Published: (2024)

Message-Passing GNNs Fail to Approximate Sparse Triangular Factorizations
by: Trifonov, Vladislav, et al.
Published: (2025)

SteeringSafety: A Systematic Safety Evaluation Framework of Representation Steering in LLMs
by: Siu, Vincent, et al.
Published: (2025)

Listener-Rewarded Thinking in VLMs for Image Preferences
by: Gambashidze, Alexander, et al.
Published: (2025)

Test-time Diverse Reasoning by Riemannian Activation Steering
by: Khanh, Ly Tran Ho, et al.
Published: (2025)

Steering Large Language Model Activations in Sparse Spaces
by: Bayat, Reza, et al.
Published: (2025)

CBMAS: Cognitive Behavioral Modeling via Activation Steering
by: Ismail, Ahmed H., et al.
Published: (2026)

HN-MVTS: HyperNetwork-based Multivariate Time Series Forecasting
by: Savchenko, Andrey, et al.
Published: (2025)

Angular Steering: Behavior Control via Rotation in Activation Space
by: Vu, Hieu M., et al.
Published: (2025)