LongPIBench: A Long-Context Benchmark for Prompt Injection
Cryptography and SecurityArtificial Intelligence
Summary
The authors point out that current tests for tricking large language models (prompt injection attacks) mostly use short pieces of text, which doesn't show the full picture. They created LongPIBench, a new test that uses very long texts from realistic tasks like reviewing papers or emails. Their tests show that existing protections don't work well when the text is very long, as simple attacks can still fool the models. They suggest LongPIBench can help researchers better evaluate and improve defenses in these real-world, long-text situations.
Authors
Yupei Liu, Yuqi Jia, Neil Zhenqiang Gong, Jinyuan Jia
Abstract
Prompt injection attacks pose a serious security risk to large language models in real-world applications. However, existing prompt injection benchmarks primarily focus on short-context inputs, leaving the attacks and defenses in long-context settings largely unexplored. This gap leads to a substantial overestimation of the effectiveness of current defenses. In this paper, we bridge the gap by introducing LongPIBench, a long-context benchmark for prompt injection covering 4 realistic application scenarios: paper peer review, resume screening, code review, and email summary. For each scenario, we construct a synthetic dataset and a real-world dataset, with context lengths ranging from thousands to tens of thousands of tokens. The evaluation results on LongPIBench reveal significant vulnerabilities of prompt injection defenses under long-context settings: even simple heuristic prompt injection attacks achieve high success rates and frequently bypass state-of-the-art defenses. We hope LongPIBench can serve as a practical benchmark for systematically evaluating prompt injection defenses in realistic long-context scenarios.