Prompt Injection Detection Dataset: 5,988 Synthetic Safe and Malicious Prompts

5,988 synthetic prompts designed for training and evaluating prompt injection detection systems. The dataset, created by Antijection, contains an approximately 50/50 split of safe and malicious examples with contextual information and attack categorization. It was last updated on January 20, 2026.

Use Cases

• Training classifiers to detect malicious prompt injections based on the described attack categorization.
• Evaluating the robustness of LLM guardrails against adversarial prompts based on the synthetic examples.
• Benchmarking different detection algorithms on a balanced set of safe and malicious prompts.
• Studying attack patterns and contextual information for prompt injection threats.

Strengths

• Contains 5,988 data points, providing a substantial corpus for model development.
• Maintains an approximately 50/50 class balance between safe and malicious prompts.
• Includes contextual information and attack categorization for each prompt.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is synthetic, which may not fully represent real-world attack patterns.

Provenance

Source

Antijection via Hugging Face.

Collection Method

Synthetically generated.

Time Range

Synthetic; no temporal coverage.

Freshness

Last updated 2026-01-20 03:41:23; freshness should be verified.

Geography

Synthetic; no spatial coverage.