Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications
Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel pith:ODALYDF2record.jsonopen to challenge →
read the original abstract
Large language models (LLMs) show inherent brittleness in their safety mechanisms, as evidenced by their susceptibility to jailbreaking and even non-malicious fine-tuning. This study explores this brittleness of safety alignment by leveraging pruning and low-rank modifications. We develop methods to identify critical regions that are vital for safety guardrails, and that are disentangled from utility-relevant regions at both the neuron and rank levels. Surprisingly, the isolated regions we find are sparse, comprising about $3\%$ at the parameter level and $2.5\%$ at the rank level. Removing these regions compromises safety without significantly impacting utility, corroborating the inherent brittleness of the model's safety mechanisms. Moreover, we show that LLMs remain vulnerable to low-cost fine-tuning attacks even when modifications to the safety-critical regions are restricted. These findings underscore the urgent need for more robust safety strategies in LLMs.
This paper has not been read by Pith yet.
Forward citations
Cited by 17 Pith papers
-
OpenSafeIntent: Evaluating Intent-Calibrated Safe Completion Across Dual-Use Prompt Sets
OpenSafeIntent benchmark shows models fail to calibrate safety across intent shifts in matched dual-use prompts, indicating current evaluations are insufficient.
-
RACC: Representation-Aware Coverage Criteria for LLM Safety Testing
RACC defines six representation-aware coverage criteria that score jailbreak test suites by measuring activation of safety concepts extracted from LLM hidden states on a calibration set.
-
Fragile Knowledge, Robust Instruction-Following: The Width Pruning Dichotomy in Llama-3.2
Width pruning in Llama-3.2 models reduces parametric knowledge while enhancing instruction-following and preserving reasoning.
-
Refusal in Language Models Is Mediated by a Single Direction
Refusal in language models is mediated by a single direction in residual stream activations that can be erased to disable safety or added to elicit refusal.
-
Faithfulness to Refusal: A Causal Audit of Neuron Selectors
A causal audit via neuron-row zeroing shows attribution methods (LRP, IG) faithfully identify dispensable neurons and can install refusal behavior, while rank-stability proxies systematically miss selector failures.
-
A Paired Testing Protocol for Batch-Conditioned Refusal Robustness in LLM Serving
The paper introduces a paired testing protocol for batch-conditioned refusal robustness in LLM serving and reports low rates of genuine safety-label flips after adjudication, with a batch-invariant kernel ablation eli...
-
Babel: Jailbreaking Safety Attention via Obfuscation Distribution Optimized Sampling
Babel is an efficient black-box jailbreaking framework that formalizes sparse safety attention heads via a mathematical obfuscation model and uses iterative distribution refinement to achieve higher attack success rat...
-
Persona-Model Collapse in Emergent Misalignment
Insecure fine-tuning raises moral susceptibility by 55% and lowers moral robustness by 65% across four frontier models, providing behavioral evidence that emergent misalignment involves persona-model collapse.
-
A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models
Suppressing one refusal neuron or amplifying one concept neuron bypasses safety alignment in LLMs from 1.7B to 70B parameters without training or prompt engineering.
-
Safety Drift After Fine-Tuning: Evidence from High-Stakes Domains
Benign fine-tuning of foundation models induces large, heterogeneous, and often contradictory changes in safety metrics across general and domain-specific benchmarks.
-
Preventing Safety Drift in Large Language Models via Coupled Weight and Activation Constraints
Coupled constraints on weight updates in a safety subspace and regularization of SAE-identified safety features preserve LLM refusal behaviors during fine-tuning better than weight-only or activation-only methods.
-
Shorter, but Still Trustworthy? An Empirical Study of Chain-of-Thought Compression
CoT compression frequently introduces trustworthiness regressions with method-specific degradation profiles; a proposed normalized efficiency score and alignment-aware DPO variant reduce length by 19.3% with smaller t...
-
REFLECTOR: Internalizing Step-wise Reflection against Indirect Jailbreak
Reflector trains LLMs to internalize step-wise self-reflection through SFT on teacher data followed by RL with outcome and validity rewards, reporting over 90% defense success against indirect jailbreaks and a 5.85% g...
-
REFLECTOR: Internalizing Step-wise Reflection against Indirect Jailbreak
Reflector internalizes step-wise self-reflection in LLMs via teacher-guided SFT then RL with outcome and validity rewards, claiming over 90% defense success against indirect jailbreaks plus utility gains like 5.85% on GSM8K.
-
Persona-Model Collapse in Emergent Misalignment
Insecure fine-tuning raises moral susceptibility 55% and lowers moral robustness 65% in four frontier models, exceeding prior benchmarks and indicating persona-model collapse as a mechanism of emergent misalignment.
-
ReGA: Model-Based Safeguard for LLMs via Representation-Guided Abstraction
ReGA uses safety-critical representations to guide abstraction in model-based analysis, enabling scalable detection of harmful LLM inputs with reported AUROC of 0.975 at prompt level.
-
Harmful Fine-tuning Attacks and Defenses for Large Language Models: A Survey
Survey of harmful fine-tuning attacks on LLMs, their variants, defense strategies, mechanical analysis, and evaluation methodologies.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.