Semantic Router: On the Feasibility of Hijacking MLLMs via a Single Adversarial Perturbation
arXiv:2511.20002v3 Announce Type: replace-cross Abstract: Multimodal Large Language Models (MLLMs) are increasingly deployed in stateless systems, such as autonomous driving and robotics. This paper investigates a novel threat: Semantic-Aware Hijacking. We explore the feasibility of hijacking multiple stateless decisions simultaneously using a single universal perturbation. We introduce the Semantic-Aware Universal Perturbation (SAUP), which acts as a semantic router, "actively" perceiving input semantics and routing them to distinct, attacker-defined targets. To achieve this, we conduct theor
The increasing deployment of MLLMs in critical, stateless systems like autonomous driving makes their vulnerabilities a pressing concern for current research.
This research reveals a critical vulnerability in MLLMs that could allow adversaries to hijack multiple decisions simultaneously with a single stealthy perturbation, posing severe security and safety risks.
The understanding of MLLM security shifts from isolated attacks to the potential for universal, semantic-aware hijacking, demanding proactive development of more robust defence mechanisms.
- · Cybersecurity firms specializing in AI
- · Developers of robust MLLM defence algorithms
- · Security researchers
- · Developers of MLLMs without robust security
- · Sectors deploying MLLMs in critical stateless systems (e.g., autonomous vehicles
- · Users reliant on MLLM-powered systems
Immediate industry-wide scramble to understand and mitigate this new class of adversarial MLLM attacks.
Increased regulatory scrutiny and demands for explainable AI and robust security measures in MLLM deployment.
Potential slowdown in the adoption of MLLMs in highly sensitive applications until adequate security standards are established.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI