| ID | Label | Category | Attack family | Context | Split | Text |
|---|
A compact test surface, not a prevalence claim.
The class ratio and family counts are deliberate design choices. They describe this release, not the frequency of attacks in production systems.
The dataset is designed around intent and trust boundaries.
Each controlled pair keeps the topic and vocabulary close while changing whether the request crosses an authorization, confidentiality or instruction-priority boundary.
What one row represents
One text input visible to an LLM application. It does not contain a model response or claim attack success.
What one pair controls
One legitimate request and one attack share a topic, source context and split without becoming paraphrases.
What the release does not claim
No production prevalence, independent annotation agreement or named-model bypass rate is claimed.
License and version
Dataset v1.0.1 is released under CC BY 4.0. Application code is released under MIT.
AltaySec affiliation · ORCID 0009-0006-9491-3565