
Securing LLM-based Agents
A new benchmark for agent security vulnerabilities and defenses
Agent Security Bench (ASB) is a comprehensive framework that formalizes and evaluates security vulnerabilities in LLM-based agents, addressing a critical gap in agent security research.
- Systematically evaluates attacks like prompt injection and memory poisoning against LLM agents
- Benchmarks effectiveness of various defense mechanisms across multiple LLM backends
- Provides a standardized methodology for assessing agent security risks
- Reveals critical insights about which defense strategies work best for different attack vectors
This research is vital for enterprises deploying LLM agents in production environments, as it establishes essential security standards and best practices for mitigating vulnerabilities in AI systems that can access tools and persist memory.
Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents