JNoether/BAD-ACTS

BAD-ACTS Dataset

Dataset Card for BAD-ACTS: Benchmark of ADversarial ACTionS

BAD-ACTS is a dataset of adversarially induced harmful actions designed to benchmark the robustness of agentic systems. It is introduced in the paper:

Benchmarking the Robustness of Agentic Systems to Adversarially-Induced Harmful Actions (2025)

This dataset accompanies the BAD-ACTS benchmark and contains examples of adversarial actions crafted to elicit harmful behavior in agentic systems across a range of task environments.

Dataset Structure

Each entry in the dataset includes:

• —Environment: The task setting (e.g., travel_planning, code_generation)
• —Category: High-level classification of the harmful behavior
• —Sub-Category: Fine-grained classification of the harmful behavior
• —Harmful_Behavior: The intended harmful action targeted by the adversary
• —Target: The agent in the environment targeted by the attack
• —Keyword: A critical keyword used for evaluation

Supported Environments

• —Travel Planning
• —Financial Article Writing
• —Code Generation
• —Multi-Agent Debate

How to Use

You can load the dataset directly using the 🤗 Datasets library:

from datasets import load_dataset

dataset = load_dataset("JNoether/bad-acts")

Dataset Size

• —Total examples: 188
• —Annotations: Human-reviewed and edited to ensure high quality and diversity

Intended Use

This dataset is intended for:

• —Evaluating the adversarial robustness of LLM agents
• —Studying failure modes and harmful behaviors
• —Training or fine-tuning safety-enhancing models
• —Developing defense mechanisms (e.g., guardian agents, prompt shields)