Indirect Prompt Injection CV Dataset for LLM Security Testing

A 2026 dataset of 27 unique Curriculum Vitae documents designed for research on indirect prompt injection in Large Language Models. Created by Alfredo Milani and colleagues, it was used to evaluate LLM robustness and threat generalizability in a controlled experimental framework. The dataset systematically varies document length and content density to isolate factors affecting model vulnerability.

Use Cases

• Benchmarking LLM robustness against indirect prompt injection based on the controlled matrix of CVs.
• Studying the impact of document length on model vulnerability as described in the experimental framework.
• Analyzing the effect of content density on security threats using the heterogeneous CV dataset.
• Developing and testing countermeasures for indirect prompt injection based on the described threat model.

Strengths

• Designed for a specific, peer-reviewed research study published in Neural Computing and Application.
• Contains a controlled matrix of 27 unique CV documents to isolate experimental variables.
• Systematically varies document length and content density as part of its experimental design.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Harvard Dataverse, authored by Alfredo Milani.

Collection Method

Designed as part of a systematic experimental framework for research.

Freshness

Last updated 2026-06-08 15:16:59; freshness should be verified.