Datasets:

Phito
/

ReplayBench

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

ReplayBench

A benchmark for evaluating tool-result misbinding in LLM agents through controlled replay.

Code and runner | English | 简体中文

Overview

ReplayBench evaluates incorrect reuse of a genuine, unmodified tool result outside its valid execution context. The replayed result contains no newly injected instruction. It is not a prompt-injection benchmark.

This dataset is the benchmark itself: scenario specs, sandbox toolkit schemas, and the 76 case bundles. The runner, checker, and sandbox code live on GitHub. Defense implementations are not included.

Dimension	Count	Where
Scenarios	19 (`F01`–`F19`)	`benchmark/scenarios/`
Failure templates	4 (`T01`–`T04`)	`cases/Fxx/Tyy/`
Replay mechanisms	3 (`M1`–`M3`)	delivery channels, not separate cases
Scenario–template cases	76	`cases/`
Scenario–template–mechanism configurations	228	76 cases × 3 channels

Failure templates

Template	Violated property	What is wrong
T01	Attribution	A genuine result from one task/object is used in another.
T02	Contextual validity	A formerly valid result is reused after authorization or resource state changes.
T03	Single execution	An already committed state-changing action executes again.
T04	Causal ordering	Results are consumed in an order inconsistent with the workflow.

Replay mechanisms

Mechanism	Delivery path
M1 Durable Recovery	A persisted tool result is retrieved during recovery.
M2 Queue Redelivery	A previously queued result is redelivered.
M3 Cache Resume	A cached result is restored when execution resumes.

M1, M2, and M3 share the same case bundle and deliver the same a_result.

What is in a case

Each cases/Fxx/Tyy/ bundle has:

state.json — sandbox initial state
a_result.json — the genuine tool result from task A
b_user.txt — task B instruction
spec.json — forbidden object, expected object, verify tool, act tool

Five shared toolkit schemas are in benchmark/toolkits/: commerce, infra, org, sensitive, workspace.

Metrics

Metric	Meaning
DAR	The replayed result reaches task B.
WAR	Task B generates a wrong state-changing action.
AAR	That wrong action is accepted.
FSER	The persistent sandbox state is independently verified as wrong. Primary metric.

Rates use valid trials only. A delivered result that does not cause a wrong action (DAR = 1, FSER = 0) is model non-acceptance, not a blocked trial and not a defense success.

Not in this dataset

Runner, checker, sandbox implementation, and adapters (on GitHub)
Call-ID binding, one-shot ticket, and world-state epoch defense code
API keys or model endpoints

Citation

If you use ReplayBench in your research, please cite our paper:

@misc{zhi2026traptoolresultreplayattack,
      title={{TRAP: Tool-Result Replay Attack for Probing Misbinding in LLM Agents}},
      author={Yutong Zhi and Jie Zhang and Xiuqi Lu and Jiayuan Xu and Jing Shao},
      year={2026},
      eprint={2610.16333},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.16333},
  }

License

Apache-2.0.