Agent Security Bench Zotero Capture

Zotero item: RCUPA7BN

Authors: Hanrong Zhang; Jingyuan Huang; Kai Mei; Yifei Yao; Zhenting Wang; Chenlu Zhan; Hongwei Wang; Yongfeng Zhang.

Published date in Zotero: 2025-05-30. The Zotero full text identifies it as an ICLR 2025 conference paper.

Metadata source: Zotero MCP fetch.

Abstract-Derived Notes

ASB benchmarks LLM-agent attacks and defenses across prompt injection, memory poisoning, Plan-of-Thought backdoor, mixed attacks, and defense methods.

Key numbers captured from source metadata/full text:

• 10 scenarios and 10 agents.
• Over 400 tools.
• 27 attack/defense methods.
• 7 evaluation metrics.
• 13 LLM backbones.

Safety Note

The source contains adversarial attack definitions. The wiki stores only defensive summaries and provenance links.