One Word at a Time: Incremental Completion Decomposition Breaks LLM Safety

Large Language Models (LLMs) are widely deployed in real-world systems. Given their broader applicability, prompt engineering has become an efficient tool for resource-scarce organizations to adopt LLMs for their own purposes. At the same time, LLMs are vulnerable to prompt-based attacks. Thus, analyzing this risk has become a critical security requirement. This work evaluates prompt-injection and jailbreak vulnerability using a large, manually curated dataset across multiple open-source LLMs, including Phi, Mistral, DeepSeek-R1, Llama 3.2, Qwen, and Gemma variants. We observe significant behavioural variation across models, including refusal responses and complete silent non-responsiveness triggered by internal safety mechanisms. Furthermore, we evaluated several lightweight, inference-time defence mechanisms that operate as filters without any retraining or GPU-intensive fine-tuning. Although these defences mitigate straightforward attacks, they are consistently bypassed by long, reasoning-heavy prompts.

Safety fine-tuning of language models typically requires a curated adversarial dataset. We take a different approach: score each candidate prompt's difficulty by how often the target model's own rollouts are judged harmful, then fine-tune on the hardest prompts paired with the model's own non-jailbroken rollouts. On Llama-3-8B-Instruct and Llama-3.2-3B-Instruct, this approach cuts the WildJailbreak attack success rate from 11.5% and 20.1% down to 1-3%, but pushes refusal on jailbreak-shaped benign prompts from 14-22% to 74-94%. Interleaving the same hard prompts 1:1 with adversarially-framed benign prompts (prompts that look like jailbreaks but have benign intent) cuts that refusal back down to 30-51% on 8B and 52-72% on 3B, at a cost of 2-6 percentage points of attack success rate. Within the mixed regime, training on the hardest half of the eligible pool rather than a random half cuts the remaining ASR by 35-50% (about 3 percentage points) on both models.

Large Language Models (LLMs) have been extensively used across diverse domains, including virtual assistants, automated code generation, and scientific research. However, they remain vulnerable to jailbreak attacks, which manipulate the models into generating harmful responses despite safety alignment. Recent studies have shown that current safety-aligned LLMs undergo shallow safety alignment. In this work, we conduct an in-depth investigation into the underlying mechanism of this phenomenon and reveal that it manifests through learned ''safety trigger tokens'' that activate the model's safety patterns when paired with the specific input. Through both analysis and empirical verification, we further demonstrate the high similarity of the safety trigger tokens across different harmful inputs. Accordingly, we propose D-STT, a simple yet effective defense algorithm that identifies and explicitly decodes safety trigger tokens of the given safety-aligned LLM to activate the model's learned safety patterns. In this process, the safety trigger is constrained to a single token, which effectively preserves model usability by introducing minimum intervention in the decoding process. Extensive experiments across diverse jailbreak attacks and benign prompts demonstrate that D-STT significantly reduces output harmfulness while preserving model usability and incurring negligible response time overhead, outperforming ten baseline methods.