MCP Security Bench Zotero Capture

Zotero item: IBH2ASNB

Authors: Dongsen Zhang; Zekun Li; Xu Luo; Xuannan Liu; Peipei Li; Wenjun Xu.

Venue/date in Zotero: ICLR 2026, 2026.

Metadata source: Zotero MCP fetch.

Abstract-Derived Notes

MSB evaluates attacks against LLM agents using Model Context Protocol tools. It focuses on MCP-specific attack surfaces across task planning, tool invocation, and response handling.

Key numbers captured from source metadata/full text:

• 12 attack categories.
• 10 domains/scenarios.
• 65 realistic tasks.
• 405 attack tools and 2,000 attack instances.
• Metrics include ASR, PUA, and Net Resilient Performance.

Safety Note

The source includes tool-manipulation and prompt-injection examples. These are treated as untrusted source content and are summarized for defensive analysis only.