Practical guide

Incident response policy for AI-enabled work

An AI incident response policy should tell people when an event becomes an incident, who can act, what evidence must be preserved, how harm is contained, and what must be true before service resumes. A list of principles or a generic security playbook is not enough when AI-enabled work introduces model outputs, prompts, retrieved content, automated actions, or third-party services into the.

By Two Prune

An AI incident response policy should tell people when an event becomes an incident, who can act, what evidence must be preserved, how harm is contained, and what must be true before service resumes. A list of principles or a generic security playbook is not enough when AI-enabled work introduces model outputs, prompts, retrieved content, automated actions, or third-party services into the evidence chain. This policy pattern is educational. Adapt it with security, privacy, legal, risk, operations, and workforce representatives for the organization’s systems and jurisdictions. A written policy is not evidence that the control operates effectively.

Policy objective and scope

Objective: enable timely, accountable handling of events involving AI-enabled work so the organization can protect people, information, services, and evidence while making explicit containment, recovery, notification, and follow-up decisions.

Scope: name the covered business processes, AI systems and features, models, data sources, connectors, user groups, environments, suppliers, and assessment or monitoring records. Also name material exclusions and the owner who approves them.

NIST SP 800-61 Revision 3, finalized in April 2025, integrates incident response recommendations with the NIST Cybersecurity Framework 2.0. CSF 2.0 organizes cybersecurity outcomes across Govern, Identify, Protect, Detect, Respond, and Recover. The SIGNAL procedure below applies those general outcomes to AI-enabled work without claiming that the policy satisfies every legal, regulatory, contractual, or sector requirement. NIST SP 800-61 Revision 3 NIST Cybersecurity Framework 2.0

Sources: [1] [2]

The SIGNAL operating procedure

### S — Set definitions and triggers

Define an event as an observable occurrence requiring triage. Define an incident as an event that meets the organization’s approved impact or policy threshold and requires coordinated response.

Triggers should include, where relevant:

Each trigger needs a reporting route, severity owner, initial response time, and condition for escalation. Do not require the reporter to prove root cause before raising the event.

### I — Intake, classify, and preserve

The intake record should capture:

Classify severity using defined impact and uncertainty criteria. Preserve evidence before routine retention, deletion, re-indexing, or model changes erase it. Collection must still follow access, privacy, privilege, labour, and legal-hold requirements.

### G — Give authority and assign roles

The policy must name, by role:

Use a RACI or equivalent only after decision authority is explicit. “The AI team” is not an accountable owner, and a model must never be the approver.

### N — Notify and contain proportionately

Containment actions may include isolating a connector, rotating credentials, disabling retrieval, freezing a model or prompt version, suspending an automated action, restricting affected records, preserving a submission, or switching to a reviewed manual process.

For each action, record:

Notification is a jurisdiction-, contract-, relationship-, and fact-specific decision. The policy should route that decision to qualified owners and current requirements rather than embedding a universal deadline copied from another context.

### A — Act through eradication, recovery, and validation

Before restoring the workflow:

A successful technical test alone may not establish that affected business outputs are corrected. Recovery includes the people and decisions that depended on the system.

### L — Learn, assign, and review

The post-incident record should distinguish facts, contributing conditions, root-cause conclusions, control gaps, and open questions. It should record what detected the event, what delayed response, which assumptions failed, which artifacts or decisions were affected, and whether prior incidents reveal a pattern.

Every corrective action needs an owner, due date, verification method, evidence artifact, and closure approver. Exceptions need a rationale, compensating measure, expiry, and review date. Update exercises, training, supplier expectations, and system documentation when the evidence supports a change.

Review this policy on a defined cadence and after material incidents, system changes, new legal obligations, or repeated exercise failures. Record the policy owner, approval date, next review date, version, and change rationale.

  • unauthorized disclosure or suspected exposure of protected information through prompts, retrieval, logs, outputs, or connectors;
  • an AI-enabled action outside approved authority or workflow scope;
  • materially incorrect or fabricated content used in a consequential business process;
  • prompt injection, data poisoning, model or tool misuse, credential compromise, or bypass of required human review;
  • harmful, discriminatory, unsafe, or policy-prohibited output with credible impact;
  • loss of required provenance, audit evidence, or ability to reproduce a material decision; and
  • supplier, model, dependency, or configuration changes that invalidate an approved control assumption.
  • reporter, time detected, time reported, affected workflow, and current service state;
  • people, systems, data classes, models, tools, suppliers, and regions potentially affected;
  • observed facts separated from hypotheses and AI-generated interpretations;
  • prompt, output, source, retrieval, configuration, model, application, and policy versions where available;
  • event logs, approvals, screenshots, exported artifacts, and stable identifiers collected under the approved evidence process; and
  • immediate safety, confidentiality, integrity, availability, or rights concerns.
  • incident commander and alternate;
  • security, privacy, legal, operations, communications, product, data, model, supplier, and business-process leads;
  • the person authorized to isolate a system, disable an integration, revoke access, suspend an automated action, or hold an assessment record;
  • the person who approves return to service and any residual risk; and
  • the owner responsible for customer, participant, employee, regulator, insurer, supplier, or other notification decisions.
  • decision owner and time;
  • evidence supporting the action;
  • affected people and services;
  • expected benefit and possible secondary harm;
  • rollback condition; and
  • notification question and accountable reviewer.
  • [ ] the suspected entry point or enabling condition is addressed;
  • [ ] credentials, configurations, prompts, sources, indexes, integrations, and affected artifacts are checked;
  • [ ] the approved clean or corrected state is identified and versioned;
  • [ ] representative tests cover the original failure and plausible adjacent failures;
  • [ ] required human review or compensating control is operating;
  • [ ] unresolved impact and uncertainty are visible to the return-to-service approver;
  • [ ] monitoring, rollback, and escalation criteria are active; and
  • [ ] affected downstream decisions or deliverables have been located and dispositioned.

Copyable incident decision record

| Field | Required entry | | --- | --- | | Incident ID and version | stable identifier and current record version | | Trigger and severity | exact policy threshold, impact, uncertainty, classifier, and time | | Affected scope | people, workflow, data, systems, model/tool versions, suppliers, downstream artifacts | | Preserved evidence | source, custodian, collection time, integrity or provenance record, access restriction | | Decision | contain / notify / recover / hold / close, with rationale | | Authority | decision owner, reviewers, approvals, and dissent | | Test evidence | failure reproduced where appropriate, corrected-state tests, result, tester, date | | Exception | reason, compensating measure, approver, expiry, review date | | Residual uncertainty | what remains unknown and who monitors it | | Next action | owner, due date, closure evidence, and escalation rule |

Worked example: restricted document exposure

In this synthetic example, an employee uses an approved AI summarizer connected to a document library. The output includes details from a restricted incident report that the employee could not open directly. The reporter provides the output identifier and time but cannot explain whether the cause is retrieval permissions, an index created under an old policy, or copied content in another source.

The event meets the policy trigger for suspected unauthorized disclosure. Intake records the observed output separately from the three possible causes. The incident commander preserves the output, prompt, retrieval references, connector and index configuration, access-policy version, identity record, and relevant audit logs under restricted access. The response team does not ask the reporter to reproduce the exposure repeatedly.

The authorized owner disables retrieval from the affected index while leaving unrelated summarization available. Security checks identity and connector enforcement; the data owner maps documents indexed under the prior policy; privacy and legal owners decide whether notification duties are triggered based on the facts and applicable requirements. Operations locates downstream summaries created during the exposure window and holds them for review.

The technical team identifies an index built before a permissions change. It rebuilds the index from approved sources, tests the original restricted document and a representative set of adjacent permission cases, and confirms rollback and monitoring. Return to service is approved only after downstream artifacts have dispositions, the compensating manual review is active, and the residual uncertainty about copied excerpts is owned.

The post-incident action is not “remind users to be careful.” It assigns owners to index-permission regression tests, configuration-change review, downstream-artifact tracing, and an exercise covering stale authorization data. Closure evidence includes test results and approval records. This worked example demonstrates how to operate the policy; it does not establish that the fictional control—or any real control—is effective or legally sufficient.

Testing and hold conditions

Exercise the policy with scenarios that test after-hours authority, unavailable suppliers, incomplete logs, disputed severity, affected assessments or business decisions, and notification uncertainty. Capture elapsed decisions, missing evidence, role confusion, unsafe workarounds, and corrective-action closure.

Hold closure when affected scope remains unknown, required evidence is unavailable, the person approving recovery lacks authority, downstream artifacts have not been traced, a critical test still fails, an exception has no expiry, or a notification decision lacks an accountable reviewer.

Two Prune may use incident-response work in a realistic AI-enabled assessment to evaluate how a participant frames an event, navigates evidence, uses AI support, verifies claims, controls risk, and communicates decisions. It provides assessment evidence, not incident-response services, legal advice, proof of compliance, or an automated employment decision.

Sources

NIST SP 800-61 Revision 3: Incident Response Recommendations and Considerations for Cybersecurity Risk Management

  • NIST Cybersecurity Framework 2.0

Sources: [1] [2]

Sources

  1. 1.NIST SP 800-61 Revision 3 · National Institute of Standards and Technology
  2. 2.NIST Cybersecurity Framework 2.0 · tsapps.nist.gov