Researchers have introduced INFUSE, a unified surface benchmark for indirect prompt injection in agentic systems, according to a new arXiv paper. The benchmark targets LLM-based agents that increasingly execute long-horizon, end-to-end workflows involving shell commands, code modification, dependency installation, and interaction with remote services.

Indirect prompt injection occurs when malicious instructions are hidden in data that an agent processes. The abstract notes that the growing autonomy of these agents expands their exposure to such attacks, though the full methodology of the benchmark is not detailed in the abstract. INFUSE aims to provide a standard evaluation surface for this emerging security challenge.