Prompt Injection Attacks with a 6-Dimensional Threat Taxonomy

NeuralAlchemy's Prompt Injection V2 Categorized dataset provides a multi-label classification system for prompt injection attacks. It supersedes a popular earlier version by moving beyond a simple binary safe/unsafe label. The dataset was last updated on June 14, 2026.

Use Cases

• Training multi-label classifiers for prompt injection detection based on the described 6-dimensional taxonomy.
• Benchmarking the robustness of LLM guardrails against categorized attack vectors.
• Analyzing the distribution and characteristics of different prompt injection threat types.
• Developing fine-grained safety evaluation suites for AI assistants and chatbots.

Strengths

• Provides a 6-dimensional taxonomy for categorizing prompt injections, which is more detailed than a simple binary label.
• Is described as the definitive successor to a highly popular previous dataset.
• Is demerged into clean, individual subsets for specific analysis.

Limitations

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

Provenance

Source

NeuralAlchemy

Freshness

Last updated 2026-06-14 09:43:58; freshness should be verified.