CyBiasBench

CyBiasBench: Benchmarking Bias in LLM Agents for Cyber-Attack Scenarios
Taein Lim, Seongyong Ju, Munhyeok Kim, Hyunjun Kim, and Hoki Kim
NeurIPS 2026 Evaluations and Datasets Track

Paper · Code and release · Interactive dashboard

A benchmark for behavioral bias in LLM offensive-security agents. Each session records one agent's interaction with an intentionally vulnerable target application, capturing not just whether attacks succeeded but how the agent chose, sequenced, and stuck with attack families.

This dataset mirrors the tabular files from the v1.0.0 GitHub release. It contains 630 sessions from five deployed agents on three web targets:

• Bias observation: 180 free-choice sessions, crossing four prompt and output-format conditions with three repetitions.
• Bias injection: 450 sessions with explicit requested-family steering, crossing ten requested attack families with three repetitions.

The agents are claude, codex, gemini, glm, and kimi; the targets are juice-shop, mlflow, and vuln-shop. Agent and target configurations are documented in the paper and the benchmark harness.

Load the data

pip install datasets

from datasets import load_dataset

repo_id = "taeng0204/CyBiasBench"
observation = load_dataset(
    repo_id, "bias_observation_records", split="test", revision="v1.0.0"
)
injection = load_dataset(
    repo_id, "bias_injection_records", split="test", revision="v1.0.0"
)
print(len(observation), len(injection))  # 180, 450

Select a configuration below to load a companion table. All configurations use the existing Parquet files; the corresponding CSV files are also included for direct download. No custom dataset loading script is required.

Configuration	Rows	Columns	Unit
bias_observation_records	180	11	Free-choice session
bias_observation_family_long	1,082	7	Session and observed attack family
bias_observation_traces	19,116	14	HTTP request
bias_injection_records	450	13	Requested-family session
bias_injection_family_long	1,597	8	Session and observed attack family
bias_injection_traces	258,019	14	HTTP request

Each configuration has one split named test, containing its complete released table. This is a loading convention, not a new train/test partition or a claim that the sessions form a held-out evaluation set. The six tables describe the same 630 sessions at different levels; their rows are not independent additional sessions.

Structure and metrics

Session tables contain attack counts, session ASR, and token use. Observation records additionally contain entropy, the most-selected family, and top-1 selection concentration. Injection records contain requested-family counts, requested-family ASR, and compliance. Companion tables join to session records through record_id.

• Observation schema
• Injection schema
• Trace schema and sanitization
• Metric glossary

The paper analyzes ten attack families. The trace classifier also includes deserialization, which is observed but never requested, and others for non-attack traffic. Injection rep values load from Parquet as integers 1, 2, and 3; the corresponding CSV values are written as 01, 02, and 03. Both formats are preserved from the release.

Read success counts at their documented level. Trace-level CSRF success flags are retained, while aggregate attack_success excludes CSRF success under the paper's conservative accounting policy. Some released ratios are rounded; use counts and the published analysis scripts to reproduce the exact estimators and session-bootstrap confidence intervals.

Collection and labels

The authors collected the sessions in isolated Docker environments against intentionally vulnerable applications. An intercepting proxy captured HTTP traffic, and a separate proxy recorded LLM API usage. An OWASP CRS-derived classifier assigned attack families; response-content heuristics and target-specific rules produced the automatic success flags.

This mirror retains the original automatic labels. It contains no human verdict sheets or human-corrected success-policy results. The collection process and annotation limitations are described in the datasheet.

The traces omit request and response bodies, headers, absolute timestamps, and verbatim agent reasoning. They retain request order, method, path with query string, response status, classifier outputs, and success evidence labels. Query strings may contain attack text; the trace schema describes the precise sanitization boundaries.

Intended use and limitations

CyBiasBench supports measurement of attack-selection bias, prompt stability, compliance with requested-family steering, and defensive evaluation of tool-using agents. See Responsible Use.

These measurements describe five deployed-agent snapshots on three web targets, not stable model rankings. Allocation reflects the model, system prompt, CLI, tools, and target interaction. Vulnerability availability, discoverability, execution difficulty, and request costs also affect family shares; the design does not identify intrinsic preference or separate these factors causally. Automatic success labels can contain errors or insufficiently supported positives.

Session-bootstrap analysis under the original automatic labels finds no improvement in aggregate mean session ASR under steering. It does not establish a reliable decrease for every agent. Decision value at larger scale, superiority to simple diversity heuristics, and generalization to unseen targets remain open.

The dataset is not intended for training or fine-tuning agents to attack real-world systems, or for claims about capability on unseen systems without further validation.

Reproduction and provenance

All twelve CSV and Parquet files are byte-identical to the GitHub release at commit af21aab5e0e476ae82048e85bec2c97e16d7db0b. release_manifest.json records their source and hashes; SHA256SUMS covers the files in this mirror.

The loader, analysis scripts, classifier, verifier, and experiment harness remain in the versioned GitHub repository. Follow its reproduction guide. The accompanying datasheet, responsible-use document, schemas, citation, and license are copied from that release; their references to code directories refer to the GitHub package.

croissant.json preserves the release's Croissant 1.0 and Responsible AI metadata, including checksums and CSV URLs pinned to GitHub v1.0.0. Hugging Face also generates separate Croissant metadata for the dataset viewer's Parquet representation.

License and citation

The dataset and documentation use CC BY-NC-ND 4.0, as stated in LICENSE. The analysis code in the GitHub repository uses MIT.

@inproceedings{lim2026cybiasbench,
  title = {CyBiasBench: Benchmarking Bias in LLM Agents for Cyber-Attack Scenarios},
  author = {Lim, Taein and Ju, Seongyong and Kim, Munhyeok and Kim, Hyunjun and Kim, Hoki},
  booktitle = {NeurIPS 2026 Evaluations and Datasets Track},
  year = {2026},
  url = {https://arxiv.org/abs/2605.07830}
}

Questions and bug reports: GitHub Issues. Corresponding authors: Hyunjun Kim (hjkim1@mju.ac.kr) and Hoki Kim (hokikim@cau.ac.kr).