AI Risk Management

Assessing and managing AI risk as a security manager: AI risk assessment, thresholds and treatment, AI threat and vulnerability management, and AI vendor and supply chain management.

5 minutes 5 Questions

AI Risk Management, within Advanced AI Security Management, is the systematic discipline of identifying, assessing, mitigating, and continuously monitoring risks that arise from the development, deployment, and operation of artificial intelligence systems. Unlike traditional IT risk, AI introduces …

Concepts covered
Continuous Monitoring of the AI Threat LandscapeAI Vendor Due Diligence and Security AssessmentAI Risk Register and Risk ReportingDenial of Service and Resource Abuse Against AILikelihood and Impact Analysis for AI RiskAdversarial Examples and Evasion AttacksBias, Fairness and Discrimination RiskVulnerability Management and Patching for AI ComponentsVendor Model Updates and Change NotificationSupply Chain Attacks on Models and ComponentsShadow AI Procurement and Unsanctioned ToolsData Poisoning and Training Data ManipulationRisk of Agentic AI and Autonomous Decision SystemsAI Risk Management Life Cycle and Enterprise RiskModel Drift and Performance Degradation RiskAI Bill of Materials and Component TransparencyInsecure Output Handling and Excessive AgencyModel Extraction, Inversion and Membership InferenceRisk of Generative AI and Large Language ModelsThreat Intelligence for AI SystemsRisk Ownership and Acceptance for AI SystemsVerifying Vendor AI Security Claims and AttestationsInherent and Residual Risk of AI SystemsHallucination and Output Reliability RiskIdentifying AI Risk ScenariosMonitoring Triggers for AI Risk ReassessmentVulnerability Scanning and Penetration Testing of AIEmbedded AI Features in SaaS and Vendor ProductsFourth-Party and Concentration Risk in AI ServicesData Usage, Training Rights and Confidentiality ClausesFoundation Model and API Provider DependenciesLegal, Reputational and Compliance Risk From AIOngoing Vendor Monitoring and Right to AuditOpen-Source Models, Datasets and Component RiskAccountability for Vendor AI Outcomes and IncidentsThe AI Threat Landscape and Threat ActorsSensitive Information Disclosure From AI SystemsSecurity Requirements in AI Contracts and SLAsAI Red Teaming and Adversarial TestingPrompt Injection and Jailbreak AttacksRisk Appetite, Tolerance and Thresholds for AIAI Impact Assessments and Regulatory ConformityThreat Modeling for AI SystemsExit Strategy and Lock-In for AI ServicesAI-Enabled Attacks: Phishing, Deepfakes and MalwareAI Risk Treatment: Mitigate, Accept, Avoid, Transfer
Test mode:
Start practice test

AAISM - AI Risk Management Example Questions

Test your knowledge of AI Risk Management

Question 1

An enterprise's customer-facing LLM assistant has been describing a warranty extension the enterprise does not offer, and customers are now submitting claims that cite it. The cause of the fabricated term has not yet been determined. Which action should the information security manager take FIRST?

Correct Answer: Filter replies that state warranty terms until fixed

Filtering replies that state warranty terms is the right first step. It is an immediate containment control that stops the ongoing harm: more customers receiving the fabricated term and filing claims the enterprise may be pressured to honor.
Tracing the term to its source in the RAG corpus is root cause analysis, which comes after containment, and the cause may not be the corpus at all. Fine-tuning on verified warranty documents is slow, costly and premature while the cause is still unknown. Raising the hallucination rating in the risk register is a documentation task that does nothing to stop the active damage.
Key concept: contain an active AI incident before investigating or remediating its root cause.

Question 2

In AI risk assessment, which of the following BEST describes a risk scenario?

Correct Answer: A plausible event in which a threat exploits an AI weakness to cause business impact

A risk scenario describes a plausible future event in which a threat actor or condition exploits an AI vulnerability, such as data poisoning or model evasion, and results in measurable business impact. This makes the risk concrete enough to analyze and treat.
A record of past incidents with root causes and residual ratings is an incident log or lessons-learned input, not a forward-looking scenario. A ranked list of models scored by inherent risk and control maturity is a risk register or prioritization output. The level of risk management is willing to pursue is risk appetite.
Key concept: risk scenario construction (threat, vulnerability, impact) in AI risk assessment.

Question 3

An organization deploys an LLM agent that triages inbound customer emails. It connects to the CRM through a service account with read, update and delete rights, but its task only requires reading records and opening tickets. During red teaming, a crafted email causes the agent to delete several customer records. Which control BEST addresses the underlying weakness the test exposed?

Correct Answer: Scope the service account to read records and create tickets

Restricting the service account to reading records and creating tickets removes the delete capability entirely, so no manipulated input can trigger record deletion. This fixes the real root cause: the agent had excessive permissions beyond its task.
Filtering emails for embedded instructions targets prompt injection, but filters are bypassable and leave the dangerous delete rights in place. Requiring analyst approval of deletes adds friction to a capability the agent should never have had. Logging and alerting on bulk deletions only detects the damage after it happens and does nothing to prevent it.
Key concept: least privilege for agent tool access, mitigating the excessive agency risk.

More AI Risk Management questions
963 questions (total)
Practice 100 questions
One session at a time, always new questions