Compliance/NIST 800-53

NIST 800-53 Rev 5 Evidence Mapping for AI Agents

NIST SP 800-53 Rev 5 is the federal security baseline. An AI agent inside a federal or FedRAMP-bound system behaves like a non-human actor that needs access enforcement, event logging, change control, and incident handling. Runtime authorization gives the AC, AU, CM, and IR control families an evidence path for AI agent workloads.

Last updated: May 20, 2026

What is NIST SP 800-53?

NIST Special Publication 800-53 Revision 5 was published by the National Institute of Standards and Technology in September 2020 (with errata releases through 2023). It defines security and privacy controls for federal information systems and organizations. For agent work, the practical mapping is to non-human subjects and service identities: software acting with delegated privileges needs access rules, audit records, change control, and incident handling. Federal agencies use 800-53 baselines under FISMA; FedRAMP inherits subsets; CMMC Level 2 inherits NIST SP 800-171, which derives from 800-53.

Why it applies to AI agents

Rev 5's access-control language covers subjects beyond human operators. An AI agent acting through service credentials should be treated as a governed non-human actor: it authenticates, requests actions, and creates audit records. The relevant AC, AU, CM, and IR controls apply when that actor can affect regulated systems or data.

The risk is that traditional 800-53 implementations were built for human accounts and deterministic services. Agents are neither. They authenticate once but make downstream tool calls. They are non-deterministic and may attempt unauthorized actions. Without runtime authorization at the tool boundary, AC-3 (Access Enforcement) may be evidenced at the API gateway without covering the action the agent takes.

Control mapping: NIST 800-53 Rev 5 to Veto features

The table maps the controls most directly engaged by agent activity. Inheritance into FedRAMP Moderate/High and CMMC Level 2 baselines is noted where relevant.

ControlRequirementVeto feature
AC-2Account management for non-organizational users (NPE accounts)Per-agent API keys scoped to project and organization; key rotation history in decision record
AC-3Access enforcement based on approved authorizationsRuntime authorization at each governed tool call; deny-by-default; allow only what policy explicitly permits
AC-6Least privilege restricts access to minimum necessaryPolicy-as-code requires explicit enumeration of tools and arguments per agent role
AC-6(9)Log use of privileged functionsDecision record distinguishes high-privilege tools; alert configuration on review-required tool use
AU-2Event logging for auditable eventsEach governed authorization decision is recorded; configurable event categories per tool
AU-3Content of audit records: what, when, where, source, outcome, identityDecision record entry includes tool name, timestamp, runtime context, agent ID, outcome, reviewer ID
AU-6Audit record review, analysis, and reportingDecision aggregates with filters; SIEM webhook for integration with Splunk, Sentinel, or Chronicle
AU-12Audit record generation by designated system componentsSDK records each governed decision; exportable decision record
CM-3Configuration change control with documented approvalPolicy-as-code in git; reviewer approval and policy validation where configured
CM-5Access restrictions for changeOwner review, branch protection, and signed-commit workflows where configured
IR-4Incident handling capability with detection, analysis, containment, eradication, recoveryPolicy rollback in one commit; emergency kill-switch policy; forensic trail from decision records
IR-5Tracking and documenting security incidentsDecision record query by date and outcome helps reconstruct incident timeline and action sequence

Evidence Veto provides

An assessor performing an Authority to Operate (ATO) review or 3PAO assessment expects artifacts they can map to each control. Veto outputs support those artifacts:

Policy artifacts (AC-3, AC-6)

YAML policy files enumerate tools, arguments, and approval rules per agent. They can be stored with version control history showing authorized changes.

Audit records (AU-2, AU-3, AU-12)

Decision record entries include agent ID, tool, arguments, policy version SHA, outcome, reviewer ID, and RFC 3339 timestamp. Exportable as JSON or CSV.

Change records (CM-3, CM-5)

Policy change history can include reviewer approvals, validation runs, and signed commits when your workflow requires them.

Incident artifacts (IR-4, IR-5)

Filtered decision record queries help reconstruct what an agent did before, during, and after an incident. Policy diffs document containment actions.

Implementation timeline

September 23, 2020NIST SP 800-53 Rev 5 published
December 10, 2020Errata release; FISMA implementing controls formalized
2023Rev 5 Patch 5 and ongoing maintenance release
OngoingFederal agencies must select baseline controls based on FIPS 199 categorization; assessment cadence per agency policy
ContinuousContinuous monitoring required under OMB A-130 and FedRAMP continuous monitoring requirements

Frequently asked questions

Is NIST SP 800-53 mandatory?
NIST SP 800-53 Rev 5 (published September 2020, updated December 10, 2020 and again in 2023) is mandatory for federal information systems under FISMA. Contractors handling Controlled Unclassified Information (CUI) inherit applicable controls through NIST SP 800-171. Private sector organizations may adopt 800-53 baselines when pursuing FedRAMP, StateRAMP, or DoD CMMC authorization.
Which 800-53 control families touch AI agents?
AC (Access Control), AU (Audit and Accountability), CM (Configuration Management), and IR (Incident Response) are the four families most directly engaged by AI agent systems. Specifically AC-3 (Access Enforcement), AC-6 (Least Privilege), AU-2 (Event Logging), AU-3 (Content of Audit Records), CM-3 (Configuration Change Control), CM-5 (Access Restrictions for Change), and IR-4 (Incident Handling) can require evidence that runtime authorization is enforcing, recording, and supporting response.
How does Veto map to AU-3 (Content of Audit Records)?
AU-3 requires audit records to contain what type of event occurred, when, where, source, outcome, and identity of any individuals or subjects associated with the event. Each Veto decision record entry contains: tool name (event type), timestamp (when), agent runtime context (where), agent ID (source), allow, deny, or approval outcome, and reviewer ID for approval flows (identity).
How does this relate to FedRAMP and CMMC?
FedRAMP Moderate baseline inherits 161 800-53 controls; FedRAMP High inherits 410. CMMC Level 2 maps to 110 NIST SP 800-171 controls, all derived from 800-53. If you are pursuing any of those authorizations, your AI agent authorization evidence can use the same control mappings as the rest of your system.

Related evidence resources

Agents are non-organizational users. They need AC, AU, CM, and IR evidence too.