KoreshieldPapers, controlled reproductions, and threat-model notes about untrusted input, retrieved context, and proposed actions. Labels distinguish evaluated work from conceptual analysis.
A middleware design placed between retrieval and inference, with a cross-document correlation mechanism for adversarial instructions distributed across retrieved chunks. The paper reports an evaluation over 5,000 adversarial scenarios.
Isaac N. Emmanuel, Teslim O. Kazeem, Victor E. EmmanuelA five-dimensional taxonomy for classifying indirect prompt injection by injection vector, operational target, persistence mechanism, enterprise context, and detection complexity.
Isaac N. Emmanuel, Teslim O. Kazeem, Victor E. EmmanuelConceptual analysis grounded in the current product architecture. These notes do not present unverified production benchmarks.
A threat-model note on provenance, cross-document composition, and the point where relevant data becomes model context.
Why a model declining one injected instruction is useful evidence, but not durable authorization or isolation.
Security inspection and business authorization answer different questions before an AI-triggered tool runs.
The harness, payload, captured runs, and limitations are public so the result can be inspected and repeated.
Read the reproduction