Intermediate
AI Incident Management Practitioner
Prepare for, detect, investigate and respond to AI attacks, misuse and harmful failures through practical exercises grounded in NIST IR 8621.
About this course
## About this free course
Build an operational AI incident-management capability for global organisations. Work through compromised agents, prompt injection, sensitive-data exposure, harmful outputs and supplier coordination, then assemble a response playbook and practise a tabletop exercise.
**Audience:** privacy, security, incident-response, AI-governance, product, procurement and risk teams. Familiarity with basic security and privacy concepts is helpful; coding is not required. Allow approximately six hours for reading, exercises and assessments. All examples are fictional and use synthetic data.
**Primary source:** [NIST IR 8621 — NIST Workshop on SI Incident Management (November 2026)](https://doi.org/10.6028/NIST.IR.8621). This is a workshop synthesis, not a mandatory standard or a settled technical definition of super intelligence. It uses SI terminology while several presentations discuss AI. This course uses AI for its operational scenarios and does not assert that these systems meet a super-intelligence threshold.
**What is original:** PrivacyBuilt supplies the workflows, worksheets, fictional scenarios, practice severity rubric and assessment questions. These are educational adaptations, not controls prescribed by NIST. Speaker recommendations and prospective NIST work are identified as such. No universal retention period, reporting deadline or automated-response mandate is inferred from the report.
**Your working outputs:** AI system inventory; incident-triage record; evidence plan; containment decision log; supplier responsibility map; recovery acceptance criteria; communications matrix; tabletop after-action report; and a 30/60/90-day improvement plan. Copy the worksheets in the lessons into your approved working environment. Do not enter real incident data, credentials or personal information into public tools or course answers.
**Assessment and certificate:** module knowledge checks and a final assessment use the existing PrivacyBuilt learning system. Complete every lesson and pass the final assessment at 80% or higher to qualify for a PrivacyBuilt certificate. This is not a NIST certification, accreditation or endorsement.
Written for a global audience by **PrivacyBuilt Editorial**. Legal applicability depends on activities, role, affected people, products, market connection and relevant law. Consult qualified counsel for specific obligations. Educational material, not legal advice. Source review: 4 November 2026.
What you will be able to do
- Distinguish attacks on AI systems from misuse, malfunction and harm-based incidents
- Map AI assets, non-human identities, tools and supplier dependencies
- Design proportionate telemetry and preserve evidence without creating new privacy or injection risks
- Triage incidents using impact, uncertainty and time sensitivity
- Select authorised containment actions and validate recovery
- Coordinate suppliers, reporting decisions and affected-person communications
- Run a synthetic tabletop and produce a prioritised improvement plan
Course content
- 1
Foundations and the two-track incident model
Interpret the workshop accurately and classify security and harm-based events.
- Understand the source and define an AI incident15 min
- Adapt the response lifecycle to AI systems15 min
4-question quiz
- 2
Readiness, asset visibility and responsibility
Map AI systems, identities, dependencies and decision rights before an incident.
- Build the AI asset and dependency inventory15 min
- Assign decision rights and prepare response capability15 min
4-question quiz
- 3
Detection, telemetry and evidence
Design useful signals and privacy-conscious evidence handling.
- Design detection signals and investigation-ready telemetry15 min
- Preserve evidence without creating a second incident15 min
4-question quiz
- 4
Triage, severity and investigation
Prioritise consequences, uncertainty and time-sensitive actions.
- Triage by harm, exposure and uncertainty15 min
- Investigate the workflow and test competing explanations15 min
4-question quiz
- 5
Containment and non-human identity controls
Restrict harmful capabilities through authorised, reversible decisions.
- Contain harm at the capability and identity boundary15 min
- Control non-human identities and defensive automation15 min
4-question quiz
- 6
Recovery, validation and continuous learning
Demonstrate safe restoration and turn findings into verified controls.
- Validate recovery before restoring trust15 min
- Turn incident findings into verified improvement15 min
4-question quiz
- 7
Reporting, communication and supplier coordination
Separate operational information sharing from applicable legal reporting duties.
- Separate incident reporting from legal notification15 min
- Coordinate providers and communicate verified facts15 min
4-question quiz
- 8
Scenario labs: attacks, misuse and harmful failures
Apply the lifecycle to prompt injection, agent compromise and secure-but-harmful outputs.
- Lab: prompt injection through retrieved documents25 min
- Lab: compromised agent identity and queued actions25 min
- Lab: technically secure but harmful AI outputs25 min
4-question quiz
- 9
Capstone: run an AI incident tabletop
Produce a complete incident record, after-action review and improvement plan.
- Capstone: facilitate a multi-step AI incident tabletop35 min
- Assemble your operational playbook and improvement plan15 min
4-question quiz
- 10
Final assessmentFinal assessment
Twenty scenario-based questions across the full course; existing 80% pass rule applies.
- Final assessment instructions and course boundaries10 min
10-question final assessment