Guardrails
Also called: input/output filters, content filters
Checks placed around a model that inspect its inputs and outputs and block, rewrite, or flag content that violates policy.
Guardrails sit on the wire: scanning user input for known jailbreak patterns, scanning retrieved content for injection markers, checking outputs for leaked secrets or disallowed content, validating that tool arguments match expectations. They can be rules, classifiers, or smaller models.
Why it matters
Useful as one layer, dangerous as the only one. Guardrails raise the cost of an attack and catch the unsophisticated cases, but injection phrasing is unbounded and determined attackers route around pattern-matchers. Pair them with structural controls (least privilege, allowlisting, egress limits) that don't depend on recognizing the attack.
Exposure map
Live counts of instances showing this pattern will appear here once the exposure map is collecting data.