CPPB: Controlled Prompt-Privacy Benchmark for LLM/VLM Agents

CPPB is the public benchmark surface for evaluating privacy propagation in AI agents, introduced in the BodhiPromptShield research. The dataset, authored by mabo1215, provides a controlled prompt manifest and template-stratified splits for train, dev, and test sets. It was last updated on Hugging Face on April 8, -2026.

Use Cases

• Benchmarking privacy leakage suppression based on controlled prompt stubs and metadata
• Training and evaluating prompt mediation models based on the template-stratified splits
• Analyzing privacy propagation patterns in AI agents based on benchmark-authored prompts

Strengths

• Provides a structured benchmark with template-stratified train/dev/test splits
• Focuses on a specific, emerging research area: privacy propagation in LLM/VLM agents
• Last update timestamp is explicitly provided: 2026-04 08 13:07:22

Limitations

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

Provenance

Source

Hugging Face, authored by mabo1215

Collection Method

Benchmark-authored prompt manifest, not raw third-party prompts or source images

Time Range

null

Freshness

Last updated 2026-04-08 13:07:22; freshness should be verified

Geography

null