gcjordi/CiberIA-PromptRisk-v1

CiberIA PromptRisk v1

CiberIA PromptRisk v1 is a balanced multilingual dataset for researching binary detection of potentially adversarial instructions targeting LLMs and AI agents.

• 3,072 records
• Catalan, Spanish and English
• 1,024 records per language
• 512 BENIGN + 512 ATTACK records per language
• 8 attack families
• 8 benign families, including hard negatives
• Pattern-grouped train/validation/test splits

This is a CiberIA Open Research artifact. It is intentionally separate from proprietary CiberIA assessment protocols and does not contain AIsecTest, CRS, ECP, private scoring rules or internal evaluation prompts.

Dataset structure

Split	Records	BENIGN	ATTACK	Share
train	2,448	1,224	1,224	79.69%
validation	312	156	156	10.16%
test	312	156	156	10.16%
Total	3,072	1,536	1,536	100%

Each language contains exactly the same number of records and the same class/family distribution.

Labels

label	class_name	Meaning
0	BENIGN	Legitimate or defensive instruction
1	ATTACK	Potentially adversarial or manipulative instruction

The first model trained from this dataset is intentionally binary. Rich taxonomy fields remain available for analysis and later multi-class or multi-label versions.

Attack families

Each family contains 192 records in total: 64 Catalan, 64 Spanish and 64 English.

attack_family	Scope
direct_prompt_injection	Direct attempts to replace or override prior instructions
indirect_prompt_injection	Malicious instructions embedded in external or retrieved content
jailbreak	Attempts to disable safeguards through personas, simulation or framing
secret_exfiltration	Attempts to expose system prompts, credentials or private context
tool_abuse	Attempts to trigger tools or irreversible actions without authorization
authority_manipulation	Unverified claims of rank, urgency or delegated authority
context_poisoning	Attempts to corrupt persistent memory, summaries or trusted context
obfuscation	Base64, homoglyph, spacing, fragmentation, typos, language mixing, markup, reversal, hidden comments and zero-width characters

Benign families and hard negatives

Each benign family also contains 192 records. Four families are marked as hard_negative=true, for a total of 768 hard negatives:

benign_family	Hard negative	Scope
security_education	yes	Defensive explanations that quote attack-like language
translation_task	yes	Translation and linguistic work on quoted adversarial strings
quotation_analysis	yes	Academic or corpus analysis without executing quoted instructions
defensive_configuration	yes	Policies, tests and controls designed to block attacks
incident_response	no	Incident summaries, containment and evidence preservation
software_development	no	Legitimate code and testing tasks using security vocabulary
ordinary_assistant	no	Standard, low-risk assistant tasks
authorized_administration	no	Scoped, reversible and explicitly approved administrative planning

Hard negatives reduce dependence on simple keywords. For example, a defensive request discussing “ignore previous instructions” remains BENIGN.

Split methodology and leakage control

The dataset uses two linked grouping fields:

• scenario_group binds the Catalan, Spanish and English versions of the same scenario.
• pattern_group binds every context generated from the same underlying instruction template.

Splits are assigned at pattern_group level. Therefore:

1. no translation pair crosses splits;
2. no scenario crosses splits;
3. no underlying template pattern crosses splits.

There are 160 pattern groups and 1,024 scenario/translation groups. For every attack or benign family, eight template patterns are assigned to train, one to validation and one to test. Context balancing produces 51/6/7 or 51/7/6 examples per language and family, alternating validation and test sizes.

This is stricter than a random row split, but it is still an in-domain synthetic evaluation. External adversarial testing remains necessary.

Record schema

Field	Type	Description
id	string	Stable record ID, e.g. CPR-CA-0001
scenario_group	string	Translation-linked scenario group
pattern_group	string	Split-isolated template-pattern group
template_id	string	Deterministic source template ID
split	string	train, validation or test
language	string	ca, es or en
text	string	Classifier input
label	int	0 BENIGN, 1 ATTACK
class_name	string	Human-readable class
category	string	Attack or benign family
attack_family	string/null	Rich attack taxonomy for ATTACK rows
benign_family	string/null	Benign taxonomy for BENIGN rows
risk_level	string	none, high or critical
obfuscated	bool	Whether obfuscation is the primary attack family
obfuscation_type	string	Obfuscation mechanism or none
agentic	bool	Whether the scenario involves an agent/tool context
scenario	string	Operational context, such as email_agent
hard_negative	bool	Benign example containing attack-like language/context
source	string	synthetic in v1
version	string	Dataset version
char_count	int	Unicode character count
text_sha256	string	SHA-256 integrity hash of text

Usage

from datasets import load_dataset

dataset = load_dataset("gcjordi/CiberIA-PromptRisk-v1")
print(dataset)
print(dataset["train"][0])

For binary training, use text as input and label as target. Do not recompute random splits: the provided splits preserve multilingual and pattern isolation.

Creation and validation

The corpus was generated deterministically from authored multilingual patterns and controlled scenario slots. Automated validation checks:

• exact row, class, language, family and split counts;
• required fields and enum values;
• ID and normalized-text uniqueness;
• translation-group completeness;
• scenario, template and pattern split isolation;
• SHA-256 text integrity;
• exact reproducibility of published Parquet files.

The source package includes the generator, validation report and a transparent TF-IDF baseline. Automated validation does not replace independent human linguistic review.

Intended uses

• Research and teaching about prompt-risk classification
• Reproducible multilingual classifier experiments
• Pre-screening or triage as one signal in defense-in-depth systems
• Analysis of false positives on defensive/hard-negative content
• Starting point for externally curated or production-specific extensions

Out-of-scope uses

• Treating BENIGN as proof that an input is safe
• Using the dataset or derived model as the sole authorization boundary
• Claiming coverage of all attack techniques, languages or deployment contexts
• Evaluating unrelated harms or factual correctness
• Deploying without external red-team data and organization-specific calibration

Limitations

• v1 is synthetic and deliberately compact.
• Only Catalan, Spanish and English are represented.
• Texts are short single-input samples, not full conversations or multimodal artifacts.
• Operational environments, tool schemas and secret formats are fictional.
• Some stylistic patterns may be easier to classify than real attacks.
• Test metrics from this corpus measure in-domain generalization, not production security effectiveness.
• Independent native-speaker and adversarial review should precede high-stakes use.

Safety and privacy

No real credentials, private prompts, personal data or live targets are included. Reserved .invalid domains and fictitious commands are used in attack examples. Handle adversarial strings as untrusted data and do not feed them to tool-enabled systems without isolation.

License

The dataset is released under CC BY 4.0. Recommended attribution:

“CiberIA PromptRisk v1” by Jordi Garcia Castillon / CiberIA Open Research, licensed under CC BY 4.0.

Code and future model weights are licensed separately under Apache-2.0.

Citation

@dataset{garcia2026ciberiapromptrisk,
  title        = {CiberIA PromptRisk v1},
  author       = {Jordi Garcia Castillon},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/gcjordi/CiberIA-PromptRisk-v1},
  version      = {1.0.0}
}