Tsun On Kwok

Tsun On Kwok, Xi Yang, Ki Sen Hung et al.

Protecting proprietary RAG databases from unauthorized redistribution is challenging: existing watermarking methods either inject fabricated relations between real entities, polluting the knowledge base with misinformation, or embed fragile lexical patterns that adversarial paraphrasing easily removes. We propose SentinelRAG, a watermarking framework that embeds style-consistent but fictitious knowledge entries into the RAG database. Our key insight is that synthetic knowledge describing fictitious entities is unlikely to be retrieved by legitimate queries, yet can be reliably triggered through targeted probes known only to the data owner. Experiments on four datasets ranging from 2.9k to 8.8M documents demonstrate that SentinelRAG achieves statistically significant detection $p < 10^{-5}$ across all tested configurations at only a 0.1% injection rate. Compared to the state-of-the-art, our method significantly reduces the false detection rate while maintaining negligible interference with legitimate user queries.

Ki Sen Hung, Xi Yang, Chang Liu et al.

A central goal of LLM alignment is to balance helpfulness with harmlessness, yet these objectives conflict when the same knowledge serves both legitimate and malicious purposes. This tension is amplified by context-sensitive alignment: we observe that domain-specific contexts (e.g., chemistry) selectively relax defenses for domain-relevant harmful knowledge, while safety-research contexts (e.g., jailbreak studies) trigger broader relaxation spanning all harm categories. To systematically exploit this vulnerability, we propose Jargon, a framework combining safety-research contexts with multi-turn adversarial interactions that achieves attack success rates exceeding 93% across seven frontier models, including GPT-5.2, Claude-4.5, and Gemini-3, substantially outperforming existing methods. Activation space analysis reveals that Jargon queries occupy an intermediate region between benign and harmful inputs, a gray zone where refusal decisions become unreliable. To mitigate this vulnerability, we design a policy-guided safeguard that steers models toward helpful yet harmless responses, and internalize this capability through alignment fine-tuning, reducing attack success rates while preserving helpfulness.