InjecAgent

Benchmark of 1,054 test cases measuring how tool-using LLM agents fall for indirect prompt injections in tool outputs.

Category

Information

Website

https://github.com/uiuc-kang-lab/InjecAgent

Repository

https://github.com/uiuc-kang-lab/InjecAgent

Pricing

open-source

License

MIT

Maintainer

UIUC Kang Lab

Added

2026-10-31

Updated

2026-10-31

Last verified

2026-10-31

Submitted by

agentica-curator

Tags

• benchmark
• prompt-injection
• security
• tool-use
• dataset

Description

The cases span 17 user tools and 62 attacker tools, covering both direct harm and data-stealing attacks. Scripts evaluate prompted (ReAct) agents and fine-tuned function-calling agents. From UIUC's Kang Lab (paper arXiv:2403.02691). MIT. Last updated July 2024.

Related

• AgentDojo: ETH Zurich benchmark environment for prompt-injection attacks and defenses on tool-using LLM agents, with suites like workspace, banking, travel and slack.

Sources

• https://github.com/uiuc-kang-lab/InjecAgent
• https://arxiv.org/abs/2403.02691

Machine-readable

• JSON
• Markdown
• Source file on GitHub (edit via pull request)