arXiv AI

The Geometry of Refusal: Why Post-Hoc Safety Is Fragile and Pretraining-Time Safety Persists

arXiv Machine Learning
Oct 3

Refusal Localizes, the Damage Relocates: Safety Layers Under Few-Sample Fine-Tuning

The paper investigates how fine‑tuning large language models with a small number of harmful examples can erode their refusal behavior, and explores whether localizing safety‑related behavior to specific layers or directions can provide robust defenses. Experiments across six checkpoints from four model families show that harmful and benign prompts remain linearly separable after attack, and that patching clean hidden states or freezing layers up to a transition depth can restore refusal. However, attackers can bypass these defenses by spreading updates or targeting singular directions, indicating that adaptive fine‑tuning can defeat localized repairs and highlighting the need for multiple defensive checks.

By Jungseob Lee, Dongyub Jude Lee, Sugyeong Eo, Seongtae Hong, Seungyoon Lee, Heuiseok Lim
arXiv Machine Learning
Aug 11

When Skills Meet Safety: Benchmarking and Characterizing the Adaptive Jailbreak Robustness of Skill-Merged LLMs

arXiv:2608. 08542v1 Announce Type: new Abstract: Model merging has become the default way to give an aligned language model new skills without retraining: a practitioner folds task vectors from math, code, or domain specialists into a safety-aligned base using task arithmetic, TIES, or DARE.

By Yu Ma, Hongli Shi, Jing Li, Xinran Xu, Weiwei Hou
arXiv AI
Aug 19

Fool's Gold: Defensive Deception Against Safety-Removal Attacks on Open-Weight Models

The paper introduces ‘Fool’s Gold’, a defensive deception technique for open‑weight language models that hardens them against safety‑removal attacks. By training decoy responses within a differentiable simulation of the attack, the method poisons the payoff of stripped refusal mechanisms, producing confident but falsified answers to hazardous requests while preserving benign behavior. Experiments on seven models (9B‑122B) show that 51‑90% of attacked‑state responses become decoys, with the defense accounting for 27‑84% of this effect, and that the defended 122B model remains within benign‑behavior budgets.

By Mark Russinovich
arXiv AI
Aug 26

ADVERSA: Measuring Multi-Turn Guardrail Degradation and Judge Reliability in Large Language Models

The paper introduces ADVERSA, an automated red‑teaming framework that evaluates large language model safety over multiple turns by tracking continuous compliance trajectories instead of binary jailbreak outcomes. Using a fine‑tuned 70B attacker model and a structured 5‑point rubric, the authors conduct controlled experiments on three frontier victim models, measuring guardrail degradation and judge reliability through a triple‑judge consensus. Results show a 26.7% jailbreak rate with most breaches occurring early, and the study documents inter‑judge agreement, attacker drift, and attacker refusals as key factors affecting safety assessment.

By Harry Owiredu-Ashley
arXiv AI
Nov 5

SafeStart: Mitigating the Overlooked Safety Cost of Active Expert Reduction in MoE LLM Serving

The paper introduces SafeStart, a technique for Mixture-of-Experts (MoE) language models that mitigates the safety degradation caused by active expert reduction during serving. By running all experts for the prompt and the first eight generated tokens before switching to a reduced set, SafeStart largely restores the model’s refusal to harmful requests while preserving most of the speedup achieved by expert reduction. Experiments on several MoE models show that SafeStart maintains high offline decode speedups (≈85% of the original) and limits the increase in harmful responses to within a few percentage points, even under jailbreak prompts.

By Erfan Baghaei Potraghloo, Saeid Shokoufa, Seyedarmin Azizi, Arya Fayyazi, Souvik Kundu, Mehdi Kamal, Massoud Pedram
arXiv Machine Learning
Sep 7

Locating and Steering Refusal Beyond Attention

The paper investigates where the ‘refusal’ behavior of language models resides across different architectures. It finds that a single direction in the residual stream governs refusal in transformers, and that the same direction—after a rigid rotation—also governs refusal in state‑space models (SSMs). By aligning these directions and applying a detector‑triggered gate, the authors demonstrate that refusal can be effectively transferred across transformer, SSM, recurrent, and hybrid architectures, showing that safety tooling can be ported by re‑estimating the direction at each architecture’s write site rather than rebuilding it from scratch.

By Preethi Carmel Bosco, Gopalakrishnan Srinivasan