AgentDojo Zotero Capture

Zotero item: ZT2SFSME

Authors: Edoardo Debenedetti; Jie Zhang; Mislav Balunovic; Luca Beurer-Kellner; Marc Fischer; Florian Tramer.

Published date in Zotero: 2024-11-24.

Metadata source: Zotero MCP fetch.

Abstract-Derived Notes

AgentDojo evaluates LLM agents that execute tools over untrusted data. It provides realistic tasks, security test cases, and attack/defense paradigms for prompt injection in dynamic tool-calling environments.

Key numbers captured from source metadata/full text:

• 97 realistic tasks.
• 629 security test cases.
• Four environments: Workspace, Slack, Banking, Travel.
• Metrics include benign utility, utility under attack, and targeted attack success rate.

Safety Note

The source contains prompt-injection examples and attack prompts. These are untrusted source content and are not operating instructions for this vault.