oi-uae/harmful-cyber-prompts

Harmful Cyber Prompts

This dataset is published strictly for research purposes. It contains deliberately harmful cybersecurity prompts collected for safety evaluation, alignment research, and red-teaming of language models. The prompts describe illegal activity (unauthorized intrusion, malware, ransomware, phishing, and related abuse). Do not use this dataset to elicit, generate, or act on harmful content. The authors do not endorse or condone any behavior described in the prompts.

Dataset Description

harmful-cyber-prompts is a categorized set of 345 harmful cybersecurity refusal-probe prompts, in the same style as AdvBench and similar harmful-prompt benchmarks. Every prompt is an explicitly-unauthorized variant — i.e. it targets third parties, systems, or data without consent — making the correct behavior for an aligned assistant an unambiguous refusal.

The dataset was built to measure per-category refusal rates in language models (e.g. before/after safety training, abliteration, or fine-tuning experiments). The category column enables a fine-grained breakdown instead of a single aggregate score.

• Language: English
• Total rows: 345
• Splits: train

Dataset Structure

Each row contains:

Column	Type	Description
prompt	string	The harmful cybersecurity prompt
category	string	One of 20 attack-category labels (see below)

Data is stored as a single parquet file in data/.

Categories

Category	Count
intrusion	20
malware	20
phishing_social_engineering	20
credential_attacks	20
network_attacks	20
web_app_attacks	20
iot_ot_ics	20
cloud_container	20
evasion_persistence	20
ransomware	15
wireless_attacks	15
mobile_malware	15
exfiltration_espionage	15
supply_chain	15
crypto_theft	15
botnet_c2	15
exploit_dev	15
privilege_escalation	15
physical_hardware	15
cyber_harassment_doxing	15

Intended Use

• Measuring and comparing refusal rates of LLMs on harmful cyber requests, with per-category granularity.
• Red-teaming and adversarial-robustness research.
• Alignment and safety-training research (e.g. constructing refusal benchmarks or evaluating safety fine-tuning).

Out of Scope

• Any operational use. The prompts must never be used as instructions to execute, plan, or facilitate cyberattacks.
• Generating harmful completions for any purpose other than internal safety evaluation.

Limitations

• Prompts are English-only and synthetic in style; they do not cover all attack techniques or real-world phrasing diversity.
• A refusal-only benchmark does not measure partial compliance, hedged answers, or multi-turn escalation.
• Category labels are single-label; some prompts could plausibly belong to multiple categories.

Research-Use-Only License

This dataset is provided for research purposes only, under the terms in the LICENSE file. By using it you agree that you will not use the dataset or its contents to plan, conduct, or facilitate any illegal activity.