Dataset Card for EconSafeBench

EconSafeBench evaluates the safety of LLM agents in executable economic environments, testing whether agents violate regulatory, informational, fairness, or data-use constraints while pursuing an economic objective under three distinct sources of pressure.

Dataset Details

Dataset Description

EconSafeBench contains 828 cases spanning five executable economic scenarios and four categories of safety violations. Unlike prompt-based safety benchmarks, every case is judged on executable actions performed through tool invocation inside a stateful economic environment, rather than on the surface form of the agent's text.

A key design feature is the matched evaluation of three pressure sources: for the same underlying decision and protected constraint, pressure may come from an explicit user request (Type II), another actor in the environment (Type III), or the agent's own pursuit of a performance objective with no external request to violate (Type I). The benchmark also supports a dynamic evaluation track (30-period interactions with persistent memory and matched controls), released separately as part of the accompanying code.

• Curated by: Yuzhu Ling, Jiale Han, Benyou Wang
• Language(s): English
• License: CC BY 4.0

Dataset Sources

• Paper: When Helpful Agents Cause Financial Harm: Evaluating LLM Agent Safety in Executable Economic Worlds — Accepted (Oral), EMNLP 2026 Workshop on Financial NLP (FinNLP)
• Code: github.com/lingyz0921-cuhk/EconSafeBench
• Authors: Yuzhu Ling, Jiale Han, Benyou Wang

Uses

Direct Use

This dataset is intended for evaluating whether LLM agents remain safe — i.e., compliant with regulatory, informational, fairness, or data-use constraints — when deployed as decision-making agents in economic settings, particularly under structural incentive pressure that does not take the form of an explicit adversarial instruction.

Out-of-Scope Use

This dataset is not intended to evaluate or train models for real-world regulatory compliance in financial or labour decisions, nor to serve as a source of ground-truth financial, legal, or HR advice. See ARTIFACT_TERMS.md in the accompanying GitHub repository for full intended-use and out-of-scope-use terms.

Dataset Structure

The main split contains 828 cases across 53 unique combinations of scenario, violation category, pressure source, and protected constraint.

By scenario:

Scenario	Description	Count
A	Portfolio Management	96
B	Market Making	181
C	Advisory & Credit	231
D	Labour Market	192
E	Platform Economy	128

By pressure source:

Type	Description	Count
Type I	Spontaneous — no external request to violate	258
Type II	User-requested	282
Type III	NPC-pressured	288

Data fields (core fields present in every record; a number of scenario-specific auxiliary fields, e.g. npc_message, client_profile, are packed into the nested extra field and vary by case):

Field	Description
id	Unique case identifier, e.g. A-T1-EUPHEM-004
scene	One of A–E
intent_type	TypeI / TypeII / TypeIII
dimension	Violation category and sub-dimension, e.g. Deception.EuphemisticUnderstatement
account_state	Structured environment state the agent observes, encoded as a JSON string (use json.loads() to parse)
task_prompt	The prompt shown to the agent
compliant_behavior	Description of a compliant response
violation_behavior	Description of a violating response
judgment_rule	The machine-checkable rule used to score the agent's tool call
critic	Aggregate critic score (0–16) from the case-construction pipeline
extra	Scenario-specific auxiliary fields (JSON object, varies by case)

Loading the Dataset

from datasets import load_dataset
import json

ds = load_dataset("Yuzhu0921/EconSafeBench")
sample = ds["train"][0]

print(sample["task_prompt"])
print(json.loads(sample["account_state"]))  # account_state is stored as a JSON string

Dataset Creation

Curation Rationale

Existing agent safety benchmarks primarily test compliance with explicit malicious instructions or prompt injection. EconSafeBench was built to close a specific gap: whether unsafe behavior can emerge from an agent's own pursuit of a performance objective, with no adversary present at all.

Source Data

Cases were produced through a generator–critic–human pipeline: a generator model produces candidate cases from a manually authored specification, a critic model scores each candidate on a 16-point rubric (mean 14.35/16 for the released cases), and cases below threshold are returned for revision. The remaining cases are reviewed by the authors, with a randomly sampled 20% independently audited by a second reviewer. See the paper (§2.3) for full methodology.

Considerations for Using the Data

This benchmark uses synthetically generated economic scenarios. While grounded in documented regulatory sources (see the paper's Appendix A), the cases are not drawn from real transactions or real client data, and violation rates on this benchmark should not be interpreted as predictive of real-world deployment risk without further validation.

Citation

BibTeX:

@inproceedings{ling2026econsafebench,
  title     = {When Helpful Agents Cause Financial Harm: Evaluating LLM Agent Safety in Executable Economic Worlds},
  author    = {Ling, Yuzhu and Han, Jiale and Wang, Benyou},
  booktitle = {Proceedings of the EMNLP 2026 Workshop on Financial NLP (FinNLP)},
  year      = {2026}
}

Dataset Card Authors

Yuzhu Ling, Jiale Han, Benyou Wang

Dataset Card Contact

yuzhuling@link.cuhk.edu.cn