Multimodal large language models (MLLMs) have shown strong image-text understanding and reasoning capabilities, but their multimodal interfaces also introduce new attack surfaces for jailbreak attacks. Recent black-box multimodal jailbreaks typically conceal harmful intent through typographic rendering, cross-modal decomposition or contextual distraction, yet their effectiveness remains constrained on strongly safety-aligned MLLMs. Our analysis suggests that unsuccessful attacks often arise from two sources: excessive visible risk that triggers refusals, and fragmented semantic cues that hinder intent reconstruction, resulting in off-target responses. To address this, we propose ARC, an audit-guided risk calibration attack framework that performs cue-preserving textual decomposition and progressive visual masking, revises residual visible risk through textual and visual audit feedback, and organizes the processed cues with scenario-grounded cross-modal reconstruction. Extensive experiments on closed-source and open-source MLLMs show that ARC consistently outperforms representative baselines, achieving average ASRs of 90.86% on closed-source models and 72.95% on open-source models. Code is available at: https://github.com/antuer/ARC. Warning: This paper contains offensive and unsafe contents generated by MLLMs.