Securing the Gatekeepers: LLM Router Vulnerabilities
First comprehensive security analysis of LLM routing systems across their entire lifecycle
This research reveals critical security gaps in LLM routing systems that dynamically select which language models to use for different tasks.
- Training vulnerabilities: Routers are susceptible to both data poisoning and backdoor attacks during training
- Inference weaknesses: Successfully demonstrates how adversaries can manipulate routers to select less capable or compromised models
- Lifecycle approach: Identifies vulnerabilities across the complete routing system lifecycle, not just isolated components
- Real-world implications: Exposes how compromised routers can significantly degrade system performance while evading detection
As organizations increasingly deploy multi-LLM architectures with routing systems, these findings highlight urgent security considerations that must be addressed to prevent potential exploitation in production environments.