The authors constructed INFUSE, a benchmark consisting of 40 realistic, long-horizon tasks spanning debugging, security, file management, system deployment, and application development.
New benchmark exposes indirect prompt injection risks in multi-step agent tasks
The INFUSE benchmark evaluates 19 LLMs across 40 realistic workflows and 56 attack scenarios, finding mean attack success rates up to 41.9% and limited protection from existing defenses.
Academic
Thuy Dung Nguyen · Jingyi Guo · Rana Seda Dubauskas · Sreepriya Damuluru · Taylor T. Johnson · Kevin Leach
Vanderbilt University
Research Digest··3 min read
The authors introduce INFUSE, an end-to-end benchmark for evaluating indirect prompt injection (IPI) in long-horizon agent workflows.
Why this paper
From Vanderbilt University
In one line
Indirect prompt injection frequently compromises long-running LLM agents, while conversational safety guards transfer poorly and instructional hardening offers stronger but uneven protection.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§