PEAR

A benchmark that measures both the utility and the attack robustness of planner-executor multi-agent systems.

PEAR
TL;DR — Planner-executor multi-agent systems get a lot done, and they are easy to hijack. PEAR measures both sides at once: how useful the system is, and how often an attacker can make it do harm.
4scenarios
84user tasks
120base attack tasks
1,680attacked user tasks

The setup

A planner agent breaks the user’s request into steps, and an executor agent carries them out with tools. This planner-executor design is practical and widely adopted, which makes its security worth studying. PEAR injects attacks into different parts of the system and checks, task by task, whether the malicious goal was achieved.

Overview of PEAR
Overview of PEAR. Center: the planner-executor system with example user tasks. Left: the attack tasks, with red arrows marking where attacks are injected. Right: utility, attack success, and efficiency metrics.

Three kinds of harm

💣

Harmful actions

Trick the system into unauthorized operations, such as taking over a user's bank account by changing the password.

🔓

Privacy leakage

Exfiltrate personal information, private Slack channels, or travel reservations.

🔋

Resource exhaustion

Burn compute and money, for example by flooding the calendar with entries or firing off a massive number of tiny transactions.

Key findings

1

The planner matters most. A weak planner hurts clean-task performance far more than a weak executor.

2

Memory helps the planner, not the executor. Adding memory to the planner substantially improves performance; adding it to the executor barely changes anything.

3

Utility and robustness trade off. Configurations with higher utility often show higher attack success rates.

4

Attacking the planner is especially effective. Its decisions flow downstream into everything the executor does.

Utility across planner-executor configurations
Utility across planner/executor pairs from the Claude, DeepSeek, GPT, and Gemini families. Most settings reach about 65% or higher; pairings with a weak planner (red labels) drop sharply.
Utility vs. harmful attack success rate
Utility vs. attack success rate for harmful-action attacks.
Utility vs. privacy attack success rate
The same for privacy-leakage attacks. Every configuration sits above the diagonal.

Citation

@inproceedings{dong2026pear,
  title     = {PEAR: Planner-Executor Agent Robustness Benchmark},
  author    = {Dong, Shen and Zhang, Mingxuan and He, Pengfei and Ma, Li and
               Thuraisingham, Bhavani and Liu, Hui and Xing, Yue},
  booktitle = {Conference of the European Chapter of the Association for Computational Linguistics (EACL)},
  year      = {2026}
}