Compiling Activation Steering into Weights via Null-Space Constraints for Stealthy Backdoors

2026-04-14

AI Summary

This work addresses the limitation of existing backdoor attacks, which often fail to sustain malicious behavior after activation due to rapid recovery of safety alignment in language models. The authors propose a novel approach that compiles activation steering vectors into weight modifications, enabling the model to exhibit harmful behaviors only when a trigger is present while preserving normal functionality and stealth on clean inputs through null-space constraints. For the first time, the backdoor objective is shifted from manipulating output tokens to steering internal representations, combined with a closed-form optimization to jointly achieve high attack success rates and strong concealment. Experiments demonstrate that the method significantly enhances persistent attack effectiveness across multiple safety-aligned large language models and jailbreaking benchmarks, without compromising safety or general performance in non-triggered settings.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Adversarial Learning & Robustness
• Search and Optimization: Learning to Search

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Safety-aligned large language models (LLMs) are increasingly deployed in real-world pipelines, yet this deployment also enlarges the supply-chain attack surface: adversaries can distribute backdoored checkpoints that behave normally under standard evaluation but jailbreak when a hidden trigger is present. Recent post-hoc weight-editing methods offer an efficient approach to injecting such backdoors by directly modifying model weights to map a trigger to an attacker-specified response. However, existing methods typically optimize a token-level mapping that forces an affirmative prefix (e.g., ``Sure''), which does not guarantee sustained harmful output -- the model may begin with apparent agreement yet revert to safety-aligned refusal within a few decoding steps. We address this reliability gap by shifting the backdoor objective from surface tokens to internal representations. We extract a steering vector that captures the difference between compliant and refusal behaviors, and compile it into a persistent weight modification that activates only when the trigger is present. To preserve stealthiness and benign utility, we impose a null-space constraint so that the injected edit remains dormant on clean inputs. The method is efficient, requiring only a small set of examples and admitting a closed-form solution. Across multiple safety-aligned LLMs and jailbreak benchmarks, our method achieves high triggered attack success while maintaining non-triggered safety and general utility.

Problem

• backdoor
• stealthiness
• reliability
• large language models
• jailbreak

Innovation

• activation steering
• null-space constraints
• stealthy backdoors
• weight editing
• jailbreak attacks