---
language:
- en
- zh
license: apache-2.0
pretty_name: ReplayBench
tags:
- benchmark
- llm-agents
- agent-security
- tool-result-replay
---
# ReplayBench

**ReplayBench: Evaluating tool-result misbinding in LLM agents through controlled replay.**

[GitHub code and executable cases](https://github.com/Phitooo/Replaybench) | [English documentation](https://github.com/Phitooo/Replaybench/blob/main/README.md) | [中文说明](https://github.com/Phitooo/Replaybench/blob/main/README.zh-CN.md)

## Overview
ReplayBench studies how genuine, unmodified tool results can become unsafe when consumed outside the execution context in which they remain valid. Controlled replay is the evaluation mechanism, not the failure definition. The benchmark covers 19 application scenarios, 4 failure templates, and 3 replay mechanisms, yielding 76 scenario-template cases and 228 scenario-template-mechanism configurations.

## Failure templates
- T01: Cross-task wrong-object attribution.
- T02: Stale-context validity after authorization or state changes.
- T03: Duplicate-effect / non-idempotent execution.
- T04: Out-of-order causal execution.

## Replay mechanisms
- M1: Durable Recovery.
- M2: Queue Redelivery.
- M3: Cache Resume.

## Metrics
DAR (delivery), WAR (wrong action), AAR (accepted wrong action), and FSER (independently verified final-state error). Rates use valid trials, and blocked or failed trials must not be miscounted as safe outcomes.

## Availability and limitations
The executable benchmark, scenario definitions, test-case bundles, replay logic, checker, example traces and paper-reported results are maintained in the linked GitHub repository. The public model-backed runner supports LlamaIndex. The paper also reports LangGraph and OpenAI Agents SDK results, but those runners are not included in the public code release. Paper-specific Call-ID binding, one-shot-ticket and world-state-epoch defense implementations are **not** released here. Do not claim the dataset is self-contained until the Hugging Face data files are actually uploaded. No production service data or API keys are required.

## Paper
Yutong Zhi, Jie Zhang, Xiuqi Lu, Jiayuan Xu, and Jing Shao. *TRAP: Tool-Result Replay Attack for Probing Misbinding in LLM Agents*. 2026.

## License
Apache-2.0.