This research reveals critical weaknesses in router-LLMs that could compromise security and privacy when dynamically selecting AI models for different tasks.

• Routing failures occur when models inaccurately assess query complexity, leading to inappropriate model selection
• Jailbreaking vulnerabilities exist in many router systems, allowing potential bypass of safety guardrails
• Privacy concerns emerge when routers fail to identify and properly handle sensitive information
• New benchmarking framework (DSC) specifically evaluates security and privacy dimensions missing from current evaluations

For security teams, this research highlights urgent needs to strengthen router-LLM defenses before widespread deployment in sensitive environments.

How Robust Are Router-LLMs? Analysis of the Fragility of LLM Routing Capabilities