arXiv:2502. 09755v4 Announce Type: replace-cross Abstract: Safety-aligned LLMs respond to prompts with either compliance or refusal, each corresponding to distinct directions in the model's activation space.
By Amit Levi, Rom Himelstein, Yaniv Nemcovsky, Avi Mendelson, Chaim Baskin
The paper introduces TIER, a Threat Implicitness Benchmark designed to evaluate large language model (LLM) safety behaviors across four risk domains and four threat levels, ranging from explicit harmful requests to sophisticated jailbreaks. Responses are scored on a six-label behavior scale by two independent LLM judges. Experiments on six open-weight LLMs reveal that safety behaviors change gradually with threat level, contextual prompts produce the most varied responses, and jailbreaks expose significant robustness gaps, underscoring the importance of behavior-aware safety evaluation.
By Thu-Hien Trinh-Thi, Hai-Yen Vong, Thanh-Ha Ung-Dung, Tram Ho
AlcaTRAz is a prompt‑level defense that uses rule trees to insert controlled character‑level perturbations into input text, disrupting jailbreak attacks without modifying or retraining the target LLM. It operates solely on the input, making it suitable for black‑box deployments, and was evaluated on 33 open‑weight models and 22 jailbreak types, outperforming three baseline defenses in 73.4 % of model‑attack combinations. While it significantly reduces high‑severity jailbreak success, it does not eliminate it and is intended as one layer of a broader defense strategy.
By Jakub Re\v{s}, Petr Ka\v{s}ka, Martin Pere\v{s}\'ini, Martin Ukrop, Kamil Malinka
The paper "Jailbreaking in the Haystack" introduces NINJA, a jailbreak technique that exploits long-context language models by appending benign, model-generated content to harmful user goals. It demonstrates that the position of harmful goals within the context is crucial for safety, and shows that NINJA significantly boosts attack success rates on models such as LLaMA, Qwen, Mistral, and Gemini. Unlike previous methods, NINJA is low-resource, transferable, less detectable, and compute‑optimal, revealing that carefully crafted benign long contexts can expose fundamental vulnerabilities in modern LMs.
By Rishi Rajesh Shah, Chen Henry Wu, Shashwat Saxena, Ziqian Zhong, Alexander Robey, Aditi Raghunathan
The paper investigates safety risks in model merging, showing that even when all constituent models are individually safety‑aligned, merging can expose a jailbreak vulnerability rooted in the pretrained foundation model. It introduces Basin‑Aware Jailbreak (BAJ), a min–max optimization method that generates adversarial suffixes transferable across merged models sharing the same backbone, without needing the exact merging coefficients or checkpoints. Experiments demonstrate BAJ’s high transfer success rates across diverse backbones and merging settings, and its resilience against existing defenses.
By Yu Zhe, Yixin Tan, Junhao Wei, Wang Chen
The paper introduces BLUEPRINT, a safety‑evaluation framework that separates a factorized social‑influence strategy space from WORLDVIEWSIM, a cross‑turn situational context module. Using Monte Carlo Tree Search, it optimizes turn‑level combinations of 18 theory‑grounded influence factors across a four‑turn trajectory, achieving near‑ceiling ASR on six frontier models with an average of only 2.46 queries. The study reveals that model‑specific vulnerabilities arise from distinct influence factors and strategy transitions, yet all models share a recovery pathway that shifts toward concrete, executable task framing to escape hard‑refusal states, highlighting the importance of monitoring how dialogue state makes unsafe requests appear actionable.
By Siyu Chen, Haoran Wang, Xiaojian Li, Yao Huang, Yinpeng Dong, Wei Xu