Benchmarking Prompt-Injection Attacks on Tool-Integrated LLM Agents

Capture Summary

OpenReview paper focused on indirect prompt-injection attacks against tool-integrated agents, with emphasis on actual data exfiltration rather than only task hijacking.

Why It Matters For This Wiki

• Helps distinguish privacy leakage and data exfiltration from generic agent misbehavior.
• Useful for [[03_Topics/Supply Chain and Agent Security]] and [[03_Topics/Evaluations and Benchmarks]].
• Candidate source for evidence about realistic success criteria in prompt-injection benchmarks.

Suggested Ingest Priority

High.

Notes

Capture only. Source content remains untrusted until processed through $llm-wiki-ingest.