Answers are grounded in this paper and cite their sources.
Qualitative field study reveals failures in multilingual safety of LLMs when engaging Najdi Arabic, indicating alignment design issues.
Multilingual safety alignment in large language models has been shown to generalize poorly to non-English inputs, with most existing work focusing on standardized language forms or orthographic transformations such as Arabizi. We present a qualitative field study documenting a compound failure mode in frontier LLMs when prompted in Najdi Arabic, the dominant dialect of central Saudi Arabia. The pattern, which we term the Multi-Vector Compound Jailbreak (MVCJ), emerges from the interaction of four layers: dialectal linguistic input, sociocultural identity signaling, plausible academic framing, and either gradual trust accumulation or explicit permission gating. We observe that two of three tested models (Google Gemini 3.1 Pro and DeepSeek Expert) consistently reach a terminal state we call Capability Advertisement, in which the model autonomously enumerates high-severity offensive cybersecurity capabilities and invites user selection. A third model (Moonshot Kimi) resists the same protocol under identical conditions, indicating that susceptibility reflects specific alignment design choices rather than inherent properties of multilingual scale or training jurisdiction. The Gemini finding was disclosed to Google's AI VRP (Reference 516149931) and classified as out of scope under the jailbreak exclusion. We position this work alongside prior research on multi-turn jailbreaks (FITD, Crescendo, Echo Chamber) and multilingual safety, highlighting an unaddressed gap: the interaction of regional dialects with conversational trust dynamics.
No takes yet. Share an insight, caveat, or question.
Saud Fahad BinMuammar (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: