AI fundamental
Instruction/data boundaries, prompt injection, source trust, least privilege, validation, approval, auditability, and residual risk.
Guiding question
Use the answer to interpret evidence and limits rather than to substitute technical vocabulary for observation.
Related units
Open the complete Friday guide for the conceptual input, demonstration, experiment, caveat, and transfer activity.
This Friday makes security concrete as a control problem. Participants experience how LLM systems can confuse untrusted content with trusted instruction, and how workflow design, source governance, permissions, and human checkpoints affect controllability.