NIST SP 800-53 Moderate

Federal security work starts with an impact rating and then follows the baseline that the rating selects. A moderate system carries a long control set and an assessment that samples it. This starter places AI traffic inside that set and provides material for the assessor.

Categorization is the step people forget. Once a rating exists, every deviation from the baseline needs a documented reason. A system owner uses this starter to describe one traffic path in that language.

The chain is short and holds no review step. Injection screening, an identity check, redaction, a marking block list, automated traffic detection and output scoring run in that order. A request marker closes the evaluation.

Redaction has no configured patterns here, so it works from the built-in identifier set alone. Add the identifiers your categorization treats as sensitive. The marking block list refuses controlled unclassified information, official use only material and sensitive but unclassified text.

Output scoring requires a combined result of 0.80 and an accuracy result of 0.80. Automated traffic detection warns without blocking, so pair it with a rate limit that you own.

Assessment procedures, control tailoring and the security plan stay with your program. Nothing here selects a baseline or writes a plan for you. The criminal justice starter uses the same baseline for a policy specific audience, and the federal cloud starters add credential and tool restrictions.