Linked CRABS Tags
CRABS-C10, CRABS-C12, CRABS-C13, CRABS-C14, CRABS-A75, CRABS-A76, CRABS-A77
Untrusted instructions enter the model via context layers (documents, tool output, memory, connector payloads) and compete with policy intent.
Evidence required: Structured context-source declarations and deterministic findings that show untrusted instruction channels or hierarchy collisions.
Decision impact: Escalates to WARN or FREEZE when paired with action channels, sensitive access, or triad conditions.
Threat model anchor for prompt/context injection.