Relational clarifications and exclusions

Clarification on AI agent security as a permissions and tooling problem, and why legitimate non-response is a security property, not a weakness.

Operational definition and functional taxonomy of AI poisoning: training, RAG, memory, pipeline, and instruction surfaces. Reading bounds to reduce confusions.

Clarification defining the doctrinal exposure audit: structured reading of surfaces that can make consumed authority drift, and thus increase interpretive risk.

Interpretive clarification prohibiting the production of plausible hypotheses when sensitive information is not explicitly published in canonical sources.

Clarification on indirect injection: a legitimate task (summary, extraction) can ingest a hostile instruction via third-party content, if the instruction/data hierarchy is not strictly bounded.

Clarification on prompt injection as authority hierarchy reversal: separation of instruction, context, and source, and bounding of surfaces where an illegitimate instruction can be consumed as authorized.