Agentic Prompt-Injection Boundary Pairs

Agentic Prompt-Injection Boundary Pairs is a synthetic English dataset containing 1,200 examples organized as 600 controlled benign/attack pairs across 50 enterprise and agentic workflow scenarios and 12 attack families. Paired rows share workflow context, assets, roles, tools, and vocabulary while differing in whether the request preserves or crosses an instruction, authorization, trust, confidentiality, tool-use, retrieval, memory, delegated-agent, or approval boundary. The release provides scenario-isolated train, validation, and test splits; JSONL and Parquet distributions; a machine-readable schema; deterministic build and validation scripts; a manual review log; and SHA-256 checksums. The content is synthetic and author-reviewed and contains no customer or private data. It is intended for prompt-injection classification, hard-negative and false-positive analysis, and agentic security regression testing. Performance on this dataset is not evidence of production security.