Skip to content
ML Academy

Introduced on day 53

Enforceable control

A safeguard decided by code outside the model, so it still holds if the model fully complies with hostile text; prompt instructions are layers, not controls.

In detail

Examples are tool allow-lists, downstream authorisation, argument validation, approval gates and sandboxes. A control that depends on the model's cooperation (an instruction, a delimiter, a classifier) lowers risk by an amount that must be measured and can change with the model, so it is reported separately.