Imagine spending millions of dollars training a state-of-the-art, closed-source large language model (LLM), carefully hardening its safety guardrails through extensive reinforcement learning. Now imagine an attacker bypassing all of those defenses using an adversarial prompt optimized on a lightweight, open-source model running on a single consumer GPU.
This is not a hypothetical security flaw; it is a foundational property of modern deep learning known as cross-model adversarial transferability. It means that an attack optimized to exploit one model can successfully compromise an entirely different, unseen model. To understand why this happens, we must look past the surface-level tokens and examine the shared mathematical and structural foundations of modern AI.
The Surrogate Model Paradigm
In an ideal security environment, a proprietary model is a complete black box. Attackers cannot see its weights, compute its gradients, or easily guess its internal mechanics. However, cross-model transferability completely breaks this isolation.
In DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models, Boxin Wang et al. demonstrated that adversarial examples crafted against small, open-source models routinely compromise massive, closed-source systems without requiring any access to their internal weights. Attackers exploit this by using a visible, open-source "surrogate" model to generate attacks. Because they have full access to the surrogate's weights, they can use gradient-driven optimization to find specific token sequences—such as adversarial suffixes—that minimize the model's safety loss.
As frameworks like ArrAttack, introduced by Linbao Li et al. in One Model Transfer to All: On Robust Jailbreak Prompts Generation against LLMs, and DSN, developed by Yukai Zhou et al. in Don't Say No: Jailbreaking LLM by Suppressing Refusal, have shown, these optimized prompts retain their efficacy when transferred to highly guarded commercial systems. The core puzzle is: why do two entirely separate models share the exact same blind spots?
Why It Works: Shared Conceptual Geometry
The primary reason adversarial attacks transfer so fluidly is that language models do not learn isolated, idiosyncratic representations of human language. Instead, independent models trained on similar data tend to converge on a shared mathematical structure.
In Uncovering Safety Risks of Large Language Models through Concept Activation Vector, Zhihao Xu et al. analyzed the internal safety directions of different LLMs. They discovered that diverse model architectures share a remarkably consistent "conceptual geometry." Because these models are all trained to predict the next token across the same massive corpus of human text, they map semantic concepts—such as "harm," "refusal," or "compliance"—into highly aligned vectors within their latent representation spaces. An adversarial prompt that successfully steers the latent vectors of a surrogate model away from a "refusal direction" will naturally tilt the latent vectors of a target model in the exact same way.
This shared geometry is deeply rooted in common pretraining paradigms. In Recent improvements of ASR models in the face of adversarial attacks, Bhiksha Raj et al. found that models sharing a common self-supervised pretraining stage remain highly susceptible to transferred attacks, even if their subsequent fine-tuning regimes differ completely. The foundational weights capture identical linguistic structures, meaning that an adversarial perturbation that warps the representation space of a local surrogate will warp the space of a remote target in a predictable manner, a concept also explored by Rishabh Maheshwary et al. in A Context Aware Approach for Generating Natural Language Attacks.
The Instruction-Following Paradox
The second driver of transferability stems from the shared nature of safety alignment itself. To make models useful, developers train them to follow complex, hierarchical instructions. Paradoxically, the better a model becomes at following instructions, the more vulnerable it becomes to universal structural hijacking.
Consider how models handle complex prompts. In Scalable and Transferable Black-Box Jailbreaks for Language Models via Persona Modulation, Rusheb Shah et al. introduced "persona modulation," a technique where an attacker automates the discovery of a persona that bypasses safety filters. The authors found that these personas transfer seamlessly across models trained via entirely different alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF) and Constitutional AI.
Similarly, in A Wolf in Sheep’s Clothing: Generalized Nested Jailbreak Prompts can Fool Large Language Models Easily, Peng Ding et al. demonstrated that structural nesting and linguistic rewriting exploit fundamental flaws in how the transformer architecture processes hierarchical instructions. Because all modern LLMs rely on similar attention mechanisms to parse context, structural tricks that obscure malicious intent fool the underlying attention layers universally.
This architectural vulnerability extends to multi-turn interactions. In Foot-In-The-Door: A Multi-turn Jailbreak for LLMs, Zixuan Weng demonstrated that forcing a model to align with its own gradually escalating toxic history creates a conversational trajectory that breaks safety alignment across completely different model families. The vulnerability is inherent to how transformers maintain state and context, making the attack highly transferable.
Cross-Modal and Agentic Universality
As machine learning systems expand to process multiple modalities and operate autonomously as agents, the surface area for transferable attacks grows proportionally.
In Vision-Language Models (VLMs), the vision encoder often acts as an unaligned backdoor to the text-aligned LLM core. In Images are Achilles' Heel of Alignment: Exploiting Visual Vulnerabilities for Jailbreaking Multimodal Large Language Models, Yifan Li et al. showed that visual jailbreak vulnerabilities exhibit high transferability across diverse multimodal architectures. Because these models rely on shared underlying training principles to align visual and textual tokens, an adversarial image optimized to trigger a harmful concept on an open-source VLM will trigger the same concept on a proprietary VLM.
Defenses designed to counteract this, such as the plug-and-play framework introduced by Yunhan Zhao et al. in BlueSuffix: Reinforced Blue Teaming for Vision-Language Models Against Jailbreak Attacks, must leverage this exact same transferability to protect closed-source models using open-source defensive layers.
When models are deployed as autonomous agents, the risk compounds. In AGENTVIGIL: Automatic Black-Box Red-teaming for Indirect Prompt Injection against LLM Agents, Zhe Ye demonstrated that indirect prompt injections optimized against one agent architecture transfer reliably to others. This reveals systemic blind spots in how agentic logic prioritizes external data over system instructions, proving that transferability is a property of the task framing and agentic logic, not just the model weights.
Turning Vulnerability into Defense
While cross-model transferability presents a daunting security challenge, it also provides a powerful mechanism for defense. If vulnerabilities transfer across architectures, robust defensive properties can transfer too.
For instance, in Vulnerability-Aware Alignment: Mitigating Uneven Forgetting in Harmful Fine-Tuning, Liang CHEN et al. showed that vulnerability labels generated from a proxy model can be used to identify and harden the "weak links" of a target model without the massive computational overhead of mapping its specific parameters. Similarly, in DataSentinel: A Game-Theoretic Detection of Prompt Injection Attacks, Yupei Liu et al. leveraged a game-theoretic framework to train a single specialized security model that successfully detects prompt injections across entirely different backend architectures.
Ultimately, cross-model adversarial transferability proves that deep learning models are not isolated islands. Because they learn from the same human language and rely on the same structural foundations, they share a unified mathematical landscape—one where both attacks and defenses can travel freely from one model to the next.