language:
- en
license: cc-by-4.0
size_categories:
- n<1K
task_categories:
- text-generation
tags:
- cybersecurity
- evaluation
- benchmark
- false-refusal
- humanwill
pretty_name: HumanWill Cybersecurity Benchmark — initial424
configs:
- config_name: default
data_files:
- split: test
path: questions.jsonl
HumanWill Cybersecurity Benchmark
Measuring harmful refusals in AI models.
424 questions for evaluating false refusal and judged answer usefulness on legitimate corporate cybersecurity tasks. For security teams selecting AI for real workflows and researchers studying when models withhold eligible assistance.
Framework and CLI on GitHub · Interactive results · Full report (PDF) · HumanWill
What is included
| Topic | Questions |
|---|---|
| Vulnerability validation | 81 |
| Source-code security | 84 |
| Binary investigation | 80 |
| Network investigation | 89 |
| Filesystem and artifact investigation | 90 |
| Total | 424 |
Question and delivery version 0.1.0. The pack/0.1.0 directory
mirrors all ten approved delivery files from the public GitHub repository byte
for byte, including prompts, supplied evidence, provenance, policy and notices.
The top-level questions.jsonl is a derived tabular view for Hugging Face browsing;
it preserves every system/user message verbatim. Original pack files are canonical.
No model responses, historical judgments or private HTTP captures are included.
Load the questions
from datasets import load_dataset
questions = load_dataset("humanwill-ai/cybersecurity-initial424", split="test")
print(len(questions)) # 424
The test split identifies evaluation use. There is no training/validation split,
no reference-answer column and no claim that these public questions are
contamination-free. expected_behavior is a case expectation, not a model score.
To run HumanWill, follow the reference benchmark quick start. The Git checkout includes the same pack. CLI and agent users share the operation API; both support saved evidence, selective retries and inspectable policies. Provider calls require your own credentials and may incur charges. Loading this dataset does not call a model. You can extend the pack with your own questions and customize domain/topic judging criteria through the framework.
Viewer fields
| Field | Meaning |
|---|---|
id |
Original case identifier; unchanged |
family, topic |
Original family code and its descriptive display name |
system_prompt, user_prompt |
Exact original messages, including supplied evidence |
expected_behavior |
Original case expectation |
criteria_json |
Original criteria encoded as a JSON string |
provenance_json |
Original provenance encoded as a JSON string |
origin_json |
Original snapshot/record hashes encoded as a JSON string |
JSON-string fields avoid changing the original nested metadata to fit a table.
VIEWER_PROVENANCE.json documents this conversion. Use the bundled manifest and
cases with the CLI; the viewer table itself is not a HumanWill pack manifest.
Construction, review and limitations
The set combines 100 original questions with 324 cases selected after Astra blocking during discovery. It is a selected challenge set, not a vendor-neutral random sample or an estimate of everyday refusal rates. Related mechanisms may recur across cases. Question construction is HumanWill AI-assisted authorship, as recorded in the original provenance.
The owner accepted the internal review of all 424 questions and the exact bundled policy on 19 September 2026, with self-declared expertise. This is not an independently verified external audit, provider approval or certification. See REVIEW_STATUS.json.
False refusal and usefulness are separate measures. An incomplete answer is not automatically a refusal; uncertainty and justified limits require contextual assessment. This assistance-focused pack does not establish harmful-compliance performance, verified task success or overall model safety.
The frozen report compares 13 model configurations on this set. Its rates exclude unclassified outcomes, retain incomplete Meta/Gemini coverage and consolidate retries. The report uses results version 0.1.1, distinct from this question-pack version. The bundled reference policy is not a promise to reproduce historical judge requests or scores. Consult the report for the exact method.
Original pack README/license prose includes historical private-access and review- preparation language. Those files are deliberately unchanged; current public access is through this mirror and the product repository. This card records the current access state without rewriting the reviewed artifacts.
License and attribution
HumanWill's licensable rights in the questions, bundled policy and pack-specific supporting materials use CC BY 4.0. The viewer transformation and this dataset card use the same terms. Credit HumanWill — https://humanwill.ai, retain notices and indicate modifications. See LICENSE.md and the preserved original pack notice. Third-party rights remain separate. The framework is separately licensed under Apache 2.0.
Suggested citation: HumanWill (2026), HumanWill Cybersecurity Benchmark, initial424, question and delivery version 0.1.0. https://humanwill.ai. CC BY 4.0.
Contribute
Use GitHub Issues for scoped contributions and the contribution guide for review and rights requirements. Practitioner feedback on eligibility and grading is welcome. Please do not post customer data, private model responses or confidential vulnerability details in public discussions.