Show HN: Deep-XPIA – Prompt injection benchmark for multi-agent AI systems
Category: security
Tags: ai-security, prompt-injection, benchmark, multi-agent, llm-security
Score: 7.8/10 (Innovation: 8, Technical: 8, Documentation: 8, Utility: 7)
Deep-XPIA is an open-source benchmark for multi-hop cross-prompt injection attacks across multi-agent AI systems, providing a taxonomy of 8 attack patterns and a harness to measure defenses against real model output. It uniquely focuses on trust boundary failures rather than delegation depth, and its live measurements have already falsified a common depth-decay hypothesis, making it a rigorous tool for AI security research.
Target audience: AI security researchers, ML engineers, and developers building multi-agent LLM systems
Repository: https://freyzo.github.io/deep-xpia/ · Python · MIT · 4 stars
View on Hacker News