Practical guide
Security control policy for AI-enabled work
A security control policy should tell people what must happen before, during, and after AI-enabled work. It should not merely list principles or declare a control effective. The GUARD method below turns a control objective into prospective rules, named responsibilities, testable records, an exception route, and a review date. Adapt it to the organization’s systems, data, risk assessment,.
A security control policy should tell people what must happen before, during, and after AI-enabled work. It should not merely list principles or declare a control effective. The GUARD method below turns a control objective into prospective rules, named responsibilities, testable records, an exception route, and a review date. Adapt it to the organization’s systems, data, risk assessment, contracts, and jurisdiction-specific obligations.
Start with a bounded control objective
Use a statement that connects the protected asset, allowed activity, and required evidence:
> People may use approved AI services for authorized work only when the data class, account, access path, output review, and retained evidence meet the rules in this policy.
That objective is narrower and more testable than “use AI securely.” It also avoids claiming that the policy itself makes a system safe or compliant.
NIST SP 800-53 Rev. 5 organizes security and privacy controls into families and separates control selection from the implementation evidence an organization needs. NIST SP 800-53A Rev. 5 provides assessment procedures for examining, interviewing, and testing controls. Together they support a crucial distinction: a written rule is design evidence; current operating records are needed to assess whether it works in practice.
Sources: [1] [2]
Write the policy with GUARD
### G — Governance and scope
Name the policy owner, approving authority, effective date, review date, covered workers, systems, data, and activities. Link to the organization’s information-classification, access-control, incident, procurement, and records policies instead of rewriting them inconsistently.
Assign responsibilities explicitly:
| Role | Required responsibility | |---|---| | Policy owner | Maintains rules, review schedule, and exception register | | Business owner | Authorizes the work purpose and confirms output-review needs | | Security/data owner | Approves services, data classes, access paths, and monitoring requirements | | User | Follows input restrictions, verifies output, and reports incidents | | Control assessor | Samples current evidence and reports design or operating gaps |
The AI governance group may coordinate these roles, but it should not silently inherit accountability owned by business, security, privacy, legal, or records functions.
### U — Use boundaries
State allowed, conditionally allowed, and prohibited uses in observable terms. A copyable rule set might say:
text Allowed: Public or internally approved content entered through an approved account for an authorized work purpose. Conditional: Restricted data only when the named data owner has approved the service, configuration, transfer, retention, and user group in writing. Prohibited: Credentials, authentication secrets, unapproved personal data, regulated records outside the approved workflow, or attempts to bypass monitoring and access controls.
Define output boundaries too. An AI response is untrusted working material until the accountable user checks material facts, calculations, sources, confidential content, and the intended audience. High-impact actions require the organization’s designated human approval; the model response is not that approval.
### A — Access, accounts, and data handling
Require approved identities, least-privilege access, managed configuration, and a current service inventory. Specify whether personal accounts, browser extensions, external connectors, file uploads, model training on submitted content, conversation retention, and cross-border processing are permitted.
For each approved service, retain:
The NIST AI RMF Core places governance across the lifecycle and connects context mapping, measurement, and risk treatment. Apply that idea here by revisiting approval when the service, model, connector, data class, or work purpose changes—not only on an annual date.
### R — Required operating procedure and records
Convert the policy into a sequence a user and reviewer can follow:
The evidence schedule should say what record demonstrates each step. Examples include an access-group export, service configuration snapshot, approval ticket, sampled task record, output-review attestation, incident record, exception decision, and periodic control-test result. Avoid indiscriminate retention of prompts or outputs when that would conflict with the organization’s data-minimization and records rules.
### D — Deviations, testing, and review
No exception should exist only in email. Use an exception record with:
text Requestor and business owner: Rule affected and requested use: Data and systems involved: Reason the standard route is insufficient: Risk assessment and compensating measures: Security/data approvals: Start, expiry, and revocation trigger: Evidence owner and closure record:
Test design and operation separately. A quarterly sample, for example, could examine whether ten active users are in approved groups, inspect the current configuration for two services, trace five restricted-data tasks to approvals, and verify that expired exceptions were removed. Record the sampling rule, test date, assessor, evidence examined, exceptions, remediation owner, and retest date.
A passing sample supports only the stated test scope and period. It does not establish legal compliance or permanent control effectiveness. Reassess sooner when a material incident occurs, a service or model changes, a new connector is enabled, a new data class enters scope, or repeated exceptions suggest the standard process no longer fits.
- service owner and business purpose;
- approved user groups and authentication method;
- permitted data classes and disabled features;
- vendor and configuration review references;
- logging, retention, deletion, and incident contacts;
- next review date and withdrawal procedure.
- Confirm the task is authorized and identify the accountable business owner.
- Classify the intended inputs before opening the AI service.
- Use the approved account, service, configuration, and connector set.
- Remove or minimize data that the task does not require.
- Review the output for material accuracy, source support, sensitive disclosure, and unsafe actions.
- Obtain any required human approval before release or operational execution.
- Retain the evidence required for the work type and report suspected exposure or control failure through the incident route.
Sources: [3]
Worked example: a policy exception for an assessment case
An assessment team wants to use an approved AI service while building a synthetic customer-operations simulation. The standard policy permits synthetic materials but blocks external connectors. The team requests a 14-day connector exception to import a controlled document set.
The business owner confirms the purpose. The data owner verifies that the set contains no production customer records or credentials. Security limits the connector to a named test workspace and three accounts, enables available logging, and sets automatic expiry. The evidence owner retains the approval, member list, configuration snapshot, hashes of the synthetic source files, and closure screenshot.
During the scheduled test, the assessor finds a fourth account added through an inherited group. The exception is paused, the inherited membership is removed, the access query is corrected, and the sample is rerun. The record says the deviation was detected and remediated; it does not describe the control as universally effective.
Publish the minimum usable policy artifact
A release-ready policy contains the control objective, scope, role table, use boundaries, operating procedure, evidence schedule, exception form, testing cadence, escalation triggers, owner, approval date, and next review date. Publish a short worker-facing procedure beside the governed policy so people can act without interpreting a long control catalogue.
Two Prune can provide assessment evidence about participant behavior in realistic AI-enabled work. It does not certify an organization’s security controls, determine legal compliance, or make employment decisions. The customer’s accountable owners must approve the policy, verify current operation, investigate exceptions, and decide what evidence is sufficient for their environment.
Sources
- 1.SP 800-53 Rev. 5 · National Institute of Standards and Technology
- 2.SP 800-53A Rev. 5 · National Institute of Standards and Technology
- 3.NIST AI RMF Core · National Institute of Standards and Technology