AgentLeak-JSON
English synthetic dataset of matched tool-response triplets for studying indirect prompt injection and fictional recovery-token disclosure.
What happened lately
AgentLeak-JSON timeline
AgentLeak-JSON provides 1,000 synthetic records, each containing safe, neutral and malicious variants of a weather-tool response. The resulting 3,000 response texts are intended for studying indirect prompt injection, detector generalization and hidden-state probes. The English card credits Saman Barahoie and Masoumeh Rezaei and declares CC BY 4.0 licensing.
The downloadable JSON uses metadata and samples at the top level and an attack_type field in each sample. Eight attack types allow a strategy-disjoint partition. All recovery tokens are fictional, and the narrow weather-response setting does not establish production security or model robustness.