
Benchmarking Bias in LLM Agents for Cyber-Attack Scenarios
A benchmark for behavioral bias in LLM offensive-security agents. Each record summarizes one agent's session against one intentionally vulnerable target application, capturing not just whether attacks succeeded but how the agent chose, sequenced, and stuck with attack families.
Status: public release accompanying the NeurIPS 2026 Evaluations & Datasets Track paper. Paper: https://arxiv.org/abs/2605.07830. Interactive result dashboard: https://trustworthyai.co.kr/CyBiasBench/. Dataset on Hugging Face: https://huggingface.co/datasets/taeng0204/CyBiasBench.
claude, codex, gemini, kimi, glmjuice-shop, mlflow, vuln-shopguided / unguided)
with output structure (structured / unstructured).The paper analyzes 10 attack families (sqli, xss, cmdi, path_traversal,
auth_bypass, idor, ssrf, csrf, file_upload, info_disclosure) chosen via
CAPEC + CWE + OWASP WSTG analysis as the families most directly testable
against the three target applications. The trace-level classifier inherits
an eleventh label (deserialization) from OWASP CRS rule family 944; it
is never a requested target and accounts for <0.1% of trace rows. See
schema/glossary.md §"Attack-family scope" for full details.
Total release: 630 aggregated records + 277,135 trace rows ≈ 48 MB.
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.
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. Alternative request-counting units preserve
between-agent separation, while changing its magnitude. Decision value at larger
scale, superiority to simple diversity heuristics, and generalization to unseen
targets remain open. See HOW_TO_REPRODUCE.md for the reproducible analyses.
git clone https://github.com/Harry24k/CyBiasBench.git
cd CyBiasBench
pip install -e . # loader: pandas + pyarrow
pip install -e '.[jsd]' # add matplotlib + scipy + numpy for re-analysis
python -c "
from cybiasbench import load_bias_observation, load_bias_injection
obs = load_bias_observation()
inj = load_bias_injection()
print(obs.wide.shape, inj.wide.shape) # (180, 11) (450, 13)
print(obs.wide.groupby('agent').session_asr.mean()) # mean session ASR per agent
"
Or load CSV / Parquet directly:
import pandas as pd
obs = pd.read_parquet("data/bias_observation_records.parquet")
The same release is also available on Hugging Face.
With datasets installed (pip install datasets):
from datasets import load_dataset
obs = load_dataset("taeng0204/CyBiasBench", "bias_observation_records",
split="test", revision="v1.0.0")
Each of the six tables is a separate configuration. The test split contains
the complete table; it does not denote a held-out partition.
Reproduction: see
HOW_TO_REPRODUCE.mdfor the two-tier reproduction guide (R1 re-analysis from the released traces; R3 full re-experiment viabenchmark/).
CyBiasBench/
├── data/ # tabular release (CSV + Parquet)
│ ├── bias_observation_records.{csv,parquet} 180 rows
│ ├── bias_observation_family_long.{csv,parquet} 1,082 rows
│ ├── bias_observation_traces.{csv,parquet} 19,116 rows (request-level)
│ ├── bias_injection_records.{csv,parquet} 450 rows
│ ├── bias_injection_family_long.{csv,parquet} 1,597 rows
│ └── bias_injection_traces.{csv,parquet} 258,019 rows (request-level)
├── schema/
│ ├── glossary.md metric definitions
│ ├── columns_bias_observation.md
│ ├── columns_bias_injection.md
│ └── columns_traces.md
├── src/cybiasbench/ python loader package
├── scripts/ R1 re-analysis + R3 post-processing
│ ├── build_records.py final_stats JSON → aggregate CSV/Parquet
│ ├── build_traces.py labeled JSONL → request-level CSV/Parquet
│ ├── verify_consistency.py sanity suite (+ --round-trip closure)
│ ├── reviewer_revisions.py bootstrap CIs + request-unit robustness
│ ├── momentum_repro.py compliance correlations + estimator comparison
│ ├── agent_identification.py random-forest agent identification
│ ├── temporal_repro.py response to failure in request order
│ ├── classify/ OWASP CRS + victim-specific classifiers
│ ├── verify/ response heuristics + session verifier
│ ├── aggregate/ final_stats.json generator
│ └── jsd_repro/ Prompt-stability JSD reproduction
│ ├── table2_jsd_robustness.py paper Table 2 (within/between-agent JSD)
│ └── prompt_profile_stability.py paper appendix per-condition JSD
├── benchmark/ R3 re-experiment harness (Docker + agents)
│ ├── run.sh one invocation = one (agent, target, prompt) session
│ ├── agents/ 6 agent CLI Dockerfiles (CLI versions pinned)
│ ├── metrics/ LiteLLM proxy + mitmproxy HTTP logger
│ ├── prompts/ guided/unguided + 10 injection prompts
│ ├── output_formats/ 4 condition templates
│ └── victims/ locally-built target images
├── HOW_TO_REPRODUCE.md R1 + R3 reproduction walkthrough
├── DATASHEET.md Gebru et al. datasheet
├── RESPONSIBLE_USE.md dual-use guidance
├── croissant.json Croissant metadata (core + RAI)
├── LICENSE CC BY-NC-ND 4.0 (data / docs)
├── LICENSE-CODE MIT (code)
└── dashboard/ Next.js source of the interactive result dashboard
The two settings here match Sections 4 and 5 of the accompanying paper:
| Setting | Paper section | Records | Design |
|---|---|---|---|
| Bias observation | §4 ("free-choice setting") | 180 | 5 agents × 3 targets × 4 conditions × 3 reps |
| Bias injection | §5 | 450 | 5 agents × 3 targets × 10 requested families × 3 reps |