Abstract
Unified Multimodal understanding and generation Models (UMMs) have demonstrated remarkable capabilities in both understanding and generation tasks. However, we identify a vulnerability arising from the generation-understanding coupling in UMMs. The attackers can use the generative function to craft an information-rich adversarial image and then leverage the understanding function to absorb it in a single pass, which we call Cross-Modal Generative Injection (CMGI). Current attack methods on malicious instructions are often limited to a single modality while also relying on prompt rewriting with semantic drift, leaving the unique vulnerabilities of UMMs unexplored. We propose STaR-Attack, the first multi-turn jailbreak attack framework that exploits unique safety weaknesses of UMMs without semantic drift. Specifically, our method defines a malicious event that is strongly correlated with the target query within a spatio-temporal context. Using the three-act narrative theory, STaR-Attack generates the pre-event and the post-event scenes while concealing the malicious event as the hidden climax. When executing the attack strategy, the opening two rounds exploit the UMM's generative ability to produce images for these scenes. Subsequently, an image-based question guessing and answering game is introduced by exploiting the understanding capability. STaR-Attack embeds the original malicious question among benign candidates, forcing the model to select and answer the most relevant one given the narrative context. Extensive experiments show that STaR-Attack consistently surpasses prior approaches, achieving up to 93.06% ASR on Gemini-2.0-Flash and surpasses the strongest prior baseline, FlipAttack. Our work uncovers a critical yet underdeveloped vulnerability and highlights the need for safety alignments in UMMs.
Introduction: The Vulnerability of Unified Multimodal Models
The rapid advancement of Unified Multimodal Models (UMMs)—architectures like Gemini and Janus-Pro that integrate both content understanding and generation—has introduced a novel security challenge. This paper identifies a critical vulnerability stemming from the tight coupling of these two functions, termed Cross-Modal Generative Injection (CMGI).
In a CMGI attack, an adversary exploits the model's own generative capabilities to create an information-rich adversarial context (such as an image). The model's understanding function then "absorbs" this context in a subsequent turn, effectively bypassing safety filters that would typically block a direct textual request for harmful information. Unlike traditional jailbreak methods that rely on prompt rewriting—which often results in semantic drift and degraded response quality—the proposed STaR-Attack framework maintains the integrity of the original malicious query by embedding it within a complex narrative structure.
Methodology: Narrative Crafting and Multi-turn Execution
STaR-Attack (Spatio-Temporal and Narrative Reasoning Attack) is a multi-turn framework that leverages narrative structures to obscure malicious intent. The framework is built upon three core components:
- Narrative Crafting: The attack utilizes a three-act narrative structure consisting of a Setup, a Climax, and a Resolution. The malicious query is defined as the "hidden climax." An external, uncensored LLM generates benign-looking text descriptions for the Setup (pre-event) and Resolution (post-event) scenes, which are causally linked to the intended malicious event.
- Multi-turn Execution Strategy:
- Turns 1 & 2: The target UMM is prompted to generate images based on the Setup and Resolution descriptions. This process "injects" the adversarial context into the model's visual memory, priming it for the final stage.
- Turn 3 (The "Guess and Answer" Game): The model is presented with the images it generated and a list of candidate questions. This list includes the original, unmodified malicious query hidden among several benign "distractor" questions. The model is then tasked with identifying and answering the question most relevant to the visual context.
- Dynamic Difficulty Mechanism: To enhance robustness, the framework can dynamically increase the number of benign distractors if the model initially refuses to answer. This forces the model's reasoning engine to rely more heavily on the established narrative context to determine the "correct" (intended) question to answer.
Experimental Evaluation
The effectiveness of STaR-Attack was evaluated using standard benchmarks for harmful instructions, including AdvBench (520 instructions) and HarmBench (400 behaviors). The attack was tested against both open-source models (e.g., Janus-Pro) and closed-source systems (e.g., Gemini-2.0-Flash).
Key results include:
- Attack Success Rate (ASR): STaR-Attack achieved a peak ASR of 93.06% on Gemini-2.0-Flash and 92.75% on Janus-Pro, significantly outperforming baselines like FlipAttack and PAIR.
- Relevant ASR (RASR): This metric assesses whether the model's response was both unsafe and directly relevant to the query. STaR-Attack demonstrated a high RASR of 71.87% on Janus-Pro, whereas the strongest baseline (PAIR) only reached 13.27%.
Discussion and Conclusion
The success of STaR-Attack highlights a fundamental architectural weakness in UMMs. By turning the model's imagination against its own safety filters, attackers can elicit high-quality, harmful responses without the semantic drift associated with previous methods. The visual context acts as a powerful "primer," making a harmful query appear relevant and benign within the established conversation state.
While the attack currently requires a multi-turn dialogue, its high success rate against state-of-the-art models underscores the urgent need for more robust multimodal safety alignments. Future research should focus on developing defense mechanisms that can detect when a model's own generated content is being used to construct an adversarial narrative.
Summary updated