AI Risk Management

AI risk management is the discipline of identifying, measuring, and treating the risks that come with enterprise-wide AI adoption — including a threat landscape that traditional application security never had to consider, and a supply chain whose accountability lines are genuinely hard to draw.

Risk assessment, thresholds, and treatment

AI risk assessment extends familiar practice with AI-specific instruments:

• Impact assessment — evaluating the potential consequences of an AI system on the organisation, on individuals, and on society. Where personal data is involved, this includes privacy impact assessments (PIAs).
• Conformity assessment — checking an AI system against applicable regulatory and standards requirements (a central mechanism in the EU AI Act for higher-risk systems).
• Risk documentation — recording identified risks, the assessments behind them, and the decisions taken.
• Thresholds — defining acceptable levels of risk, aligned to the enterprise risk appetite set during governance.
• Treatment plans — deciding to mitigate, transfer, accept, or avoid each risk, and tracking exposure with KRIs and KPIs over time.

The AI threat landscape and vulnerability management

AI systems introduce a class of vulnerabilities that live in the model and its data, not just the code. Managing them combines testing with an up-to-date view of adversary behaviour:

• Testing — penetration testing of the surrounding infrastructure, vulnerability testing of models and pipelines, and red teaming aimed at model-level and prompt-level attacks.
• AI-specific vulnerabilities — prompt injection, model inversion, membership inference, model evasion, and training-data extraction. The OWASP Top 10 for LLM Applications and MITRE ATLAS catalogue these in depth.
• Adversarial threats — deliberate manipulation of inputs or training data to force incorrect behaviour.
• Threat intelligence and attack chains — sourcing intelligence on AI-specific threats, and understanding how adversaries use AI to enhance their own operations.
• Anomaly detection — spotting abnormal model behaviour or usage.
• Emerging vectors — deep fakes used for fraud and disinformation, insider misuse of AI, and the new risks of autonomous, tool-using AI agents.

Maintaining a current picture of this landscape is an ongoing activity, not a one-time exercise.

Vendor and supply chain risk

Most enterprises consume AI they did not build, so much of the risk lives in the supply chain:

• Software and library dependencies — the open-source and third-party components in AI stacks.
• Vendor due diligence and contracts, including SLAs, restrictions on how vendors may use enterprise data, and clear access controls.
• Accountability models — notably the provider vs. deployer distinction (whoever builds an AI system versus whoever puts it into use carries different obligations, a key EU AI Act concept), and the reality of third-, fourth-, and fifth-party risk cascading down the chain.
• Ownership and intellectual property — who owns the model, its outputs, and the training data — and how liability is allocated when AI causes harm.
• Ongoing vendor monitoring for changes in risk, rather than a point-in-time assessment.

Related pages

• AI Governance & Program Management
• AI Technologies & Controls
• Supply Chain Security · Penetration Testing
• Frontier AI & AI-Powered Threats

References

• MITRE ATLAS — Adversarial Threat Landscape for AI Systems
• OWASP Top 10 for LLM Applications
• ISO/IEC 23894 — AI risk management
• EU AI Act