PromptShield: A Benchmark for Prompt Injection Attack Detection

PromptShield is a benchmark dataset for binary classification of prompt injection attacks, accompanying the paper "PromptShield: Deployable Detection for Prompt Injection Attacks". The dataset was created by author hendzh and last updated on April 7, 2025. It is built from a curated selection of open-source datasets and published attack strategies.

Use Cases

• Train binary classifiers for prompt injection detection based on the described prompt text field.
• Benchmark the performance of security models against realistic attack strategies mentioned in the description.
• Analyze patterns in malicious prompt construction based on the curated open-source attack data.

Strengths

• Dataset is explicitly designed as a flexible and realistic benchmark for a specific security task.
• Based on a curated selection of open-source datasets and published attack strategies.
• Last updated on 2025-04-07, indicating recent maintenance.

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

hendzh on Hugging Face, accompanying the academic paper "PromptShield: Deployable Detection for Prompt Injection Attacks".

Collection Method

Built from a curated selection of open-source datasets and published prompt injection attack strategies.

Freshness

Last updated 2025-04-07 17:15:28.