Emerging AI Risk and Horizon Scanning

1. Why Emerging AI Risk and Horizon Scanning Matters

Artificial intelligence is one of the fastest-moving technology domains in history. New model architectures, capabilities, attack techniques, regulations and societal expectations appear in months rather than years. A risk register that was accurate at the start of a fiscal year may be dangerously incomplete by the end of it. This is why the ISACA Advanced in AI Risk (AAIR) body of knowledge treats emerging AI risk identification and horizon scanning as a core element of the AI Risk Program Management domain.

Traditional enterprise risk management (ERM) is largely backward-looking: it relies on loss history, incident data and known control failures. Emerging AI risks, by definition, have little or no loss history. If an organization waits until a risk has materialized to add it to the register, it has already lost the opportunity to prepare. Horizon scanning gives the risk function a forward-looking capability so that the AI risk program can anticipate rather than merely react.

For the exam, remember the key value propositions:
- Protects the organization from being blindsided by novel AI threats, regulatory changes and reputational shifts.
- Supports risk appetite and strategy discussions at board level with credible foresight.
- Enables proactive control design and budgeting rather than emergency remediation.
- Demonstrates due diligence and good governance to regulators, auditors and stakeholders.
- Keeps the AI risk register, key risk indicators (KRIs) and risk scenarios current and relevant.

2. What Is Emerging AI Risk?

An emerging risk is a risk that is new, evolving, or not yet fully understood, where the likelihood, impact or velocity is highly uncertain and where historical data is scarce or absent. In the AI context, emerging risks typically arise from one or more of the following drivers:

- Technological change: new model types (multimodal, agentic, reasoning models), new deployment patterns (autonomous agents, AI-to-AI interactions), and new attack surfaces (prompt injection, model extraction, data poisoning, jailbreaks).
- Regulatory and legal change: the EU AI Act and its phased obligations, sector-specific guidance from financial, health and consumer regulators, evolving liability doctrines, copyright and intellectual property rulings concerning training data and outputs.
- Market and supply chain change: dependence on a small number of foundation model providers (concentration risk), sudden deprecation of models or APIs, pricing changes, and open-source model proliferation.
- Societal and ethical change: shifting public tolerance for automated decision-making, deepfake-enabled fraud and misinformation, workforce displacement concerns, and expectations around transparency and explainability.
- Geopolitical change: export controls on AI chips, data localization requirements, national AI strategies and sanctions affecting vendors.

Common examples that exam questions may reference include:
- Agentic AI systems taking autonomous actions with real-world consequences.
- Model drift and model collapse caused by training on AI-generated data.
- Shadow AI: employees using unsanctioned AI tools, exposing data and bypassing controls.
- Synthetic identity and deepfake-enabled social engineering.
- Quantum computing threats to cryptography protecting AI pipelines.
- Environmental and sustainability risk tied to compute consumption.
- Third-party and fourth-party AI embedded in SaaS products without the organization's knowledge.

3. What Is Horizon Scanning?

Horizon scanning is a systematic, structured process for detecting early signs (often called weak signals) of potential threats, opportunities and developments that could affect the organization in the future. It is a form of strategic foresight. Unlike risk assessment, which evaluates known risks, horizon scanning aims to discover what is not yet on the register.

Key characteristics:
- Continuous or periodic: not a one-time exercise; it has a defined cadence (e.g., monthly scanning, quarterly synthesis, annual deep dive).
- Multi-source: draws from research publications, regulator announcements, vendor roadmaps, incident databases, threat intelligence, industry consortia, academic conferences, patent filings, media and internal signals.
- Cross-functional: involves legal, compliance, security, data science, procurement, HR, and business lines, because AI risk cuts across all of them.
- Structured output: produces documented findings that feed into the risk register, risk scenarios, KRIs, control roadmaps and board reporting.

4. How Horizon Scanning Works: The Process

A mature horizon scanning capability generally follows these steps. Exam questions frequently test the order and the purpose of each step.

Step 1 - Define scope and objectives. Clarify what the scan covers (technology, regulation, threats, markets, ethics), the time horizon (short term 0-12 months, medium 1-3 years, long term 3+ years), and the link to enterprise strategy and risk appetite.

Step 2 - Identify and curate sources. Establish a source list and weighting. Credible sources include standards bodies (ISO/IEC 42001, ISO/IEC 23894, NIST AI RMF updates), regulators, CERTs, the OWASP Top 10 for LLM Applications, MITRE ATLAS, AI incident databases, peer-reviewed research, and vendor security advisories. Internal sources include audit findings, incident tickets, employee reports and model monitoring telemetry.

Step 3 - Collect signals. Gather raw information systematically. Automation (news aggregators, regulatory trackers, threat intelligence feeds, even AI-assisted summarization) can increase coverage, but human judgment remains essential.

Step 4 - Filter and analyze. Separate noise from meaningful signals. Techniques include:
- PESTLE analysis (Political, Economic, Social, Technological, Legal, Environmental) to categorize drivers.
- Trend analysis to identify direction and velocity of change.
- Scenario planning to build plausible futures and stress-test strategy and controls.
- Delphi method and expert panels to reach structured consensus under uncertainty.
- Risk radar or heat map to plot emerging risks by proximity and potential impact.
- Cross-impact analysis to understand how one emerging risk may amplify another.

Step 5 - Assess and prioritize. Because historical data is lacking, assessment relies on qualitative scales: likelihood, impact, velocity (how fast the risk could materialize), proximity (how near in time), uncertainty, and preparedness. Risks are mapped against risk appetite and tolerance.

Step 6 - Respond and integrate. Outcomes may include: adding a new entry or scenario to the AI risk register; adjusting existing risk ratings; defining new KRIs or leading indicators; commissioning a deep-dive assessment; updating policies, controls, training or vendor contracts; or deliberately accepting and monitoring the risk. Integration with the broader risk program is what distinguishes mature horizon scanning from mere news reading.

Step 7 - Communicate and report. Findings are summarized for the AI governance committee, risk committee and board. Good reporting is concise, prioritized, action-oriented and explicit about uncertainty.

Step 8 - Review and improve. Periodically evaluate the process itself: Did we miss signals that later became incidents? Were sources adequate? Was the cadence appropriate? Lessons learned feed back into Step 1.

5. Governance Roles and Accountability

- Board and senior management: set risk appetite, demand forward-looking reporting, approve resources.
- Chief Risk Officer or AI risk lead: owns the horizon scanning process and ensures integration into ERM.
- AI governance committee: reviews emerging risk findings, decides on response.
- First line (business and data science teams): contribute operational signals and implement controls.
- Second line (risk and compliance): run the scanning process, challenge assumptions, maintain the register.
- Third line (internal audit): independently assures that horizon scanning exists, is effective and that outputs are acted upon.

6. Linking Horizon Scanning to Frameworks

Exam questions often expect you to connect horizon scanning to recognized frameworks:
- NIST AI RMF: the MAP function (understanding context and identifying risks) and the GOVERN function (ongoing processes to identify and manage emerging risks) directly support horizon scanning. MANAGE includes prioritizing and responding to risks over time.
- ISO/IEC 42001: requires understanding the organization and its context, determining risks and opportunities, and continual improvement, all of which imply monitoring the external environment.
- ISO/IEC 23894 and ISO 31000: emphasize that risk management is dynamic, iterative and responsive to change.
- COBIT: aligns risk management with strategy and emphasizes monitoring and evaluation.
- EU AI Act: phased obligations and evolving harmonized standards make regulatory horizon scanning essential for compliance planning.

7. Key Risk Indicators and Early Warning

Horizon scanning should produce leading indicators rather than only lagging ones. Examples:
- Number of new AI regulations or guidance documents issued in operating jurisdictions.
- Count of publicly reported incidents involving the model types or vendors the organization uses.
- Growth in unsanctioned AI tool usage detected via network or SaaS discovery.
- Vendor announcements of model deprecation or terms-of-service changes.
- Frequency of new jailbreak or prompt injection techniques published against models in production.
- Drift metrics approaching thresholds in deployed models.
Thresholds should be tied to risk tolerance so that breaches trigger defined escalation.

8. Common Pitfalls

- Treating horizon scanning as informal reading with no documentation or follow-through.
- Over-reliance on a single source or on vendor marketing.
- Focusing only on technology and ignoring legal, ethical and societal drivers.
- Failing to connect findings to the risk register, scenarios and KRIs.
- Paralysis by uncertainty: refusing to act because data is incomplete. Emerging risk management accepts uncertainty and uses qualitative judgment.
- Confirmation bias and groupthink in expert panels.
- Scanning too infrequently for the pace of AI change.
- Not auditing the process or learning from missed signals.

9. Exam Tips: Answering Questions on Emerging AI Risk and Horizon Scanning

Tip 1 - Recognize the forward-looking cue. If a question mentions new, evolving, future, anticipate, weak signals, trends, not yet experienced, or limited historical data, the answer almost always involves horizon scanning, scenario analysis or emerging risk processes rather than loss-data analysis or incident response.

Tip 2 - Choose systematic over ad hoc. ISACA favors structured, documented, repeatable processes. Between an option describing informal monitoring by individuals and one describing a formal, cross-functional scanning process with defined sources and cadence, choose the formal one.

Tip 3 - Integration is the goal. The BEST or MOST important outcome of horizon scanning is updating the risk register, risk scenarios, KRIs, risk appetite discussions and control plans. An option that stops at producing a report without integration is weaker.

Tip 4 - Think in terms of lines of defense. Second line typically owns the process; first line contributes signals and implements controls; third line (audit) provides assurance. Questions asking who should independently assess the effectiveness of horizon scanning point to internal audit.

Tip 5 - Board communication should be prioritized and decision-useful. If asked what to report to the board about emerging AI risks, prefer options emphasizing prioritized risks, potential strategic impact, alignment to risk appetite and proposed responses over exhaustive technical detail.

Tip 6 - Qualitative methods are legitimate for emerging risk. When historical data is unavailable, scenario analysis, expert judgment, Delphi and qualitative scales are appropriate. Do not choose answers that insist on quantitative loss data before any action.

Tip 7 - Multi-source and multi-dimensional. Correct answers usually reference diverse sources and consider regulatory, legal, ethical and societal dimensions, not just technical threats.

Tip 8 - Frequency matches velocity. If a question highlights rapid change in AI, the correct answer favors continuous or more frequent scanning and review cycles rather than annual reviews.

Tip 9 - Watch for the FIRST step. When asked what to do FIRST after identifying an emerging AI risk, the answer is typically to assess it (understand likelihood, impact, velocity, relevance to the organization) and determine whether it falls within risk appetite, before designing controls or escalating. Conversely, if the scenario says the risk is clearly beyond tolerance and imminent, escalation and response take priority.

Tip 10 - Eliminate distractors. Distractor options often: describe reactive activities (post-incident review) when the question asks about anticipation; assign the task to an inappropriate role (e.g., having the vendor define the organization's emerging risks); or propose ignoring a risk because it has not yet occurred.

Tip 11 - Map to frameworks when possible. Associate horizon scanning with NIST AI RMF MAP and GOVERN, ISO/IEC 42001 context and continual improvement, ISO 31000 dynamic risk management, and regulatory change management for the EU AI Act.

10. Worked Example

Scenario: An organization relies heavily on a third-party foundation model. The vendor announces a major architecture change and new usage terms effective in six months. The AI risk manager learned this from a news article. What is the BEST course of action?

A. Wait for the vendor to provide formal notice before acting.
B. Log the signal, assess the potential impact on existing use cases against risk appetite, update the AI risk register and KRIs, and escalate to the AI governance committee with options.
C. Immediately terminate the vendor relationship.
D. Ask internal audit to perform a vendor audit.

Analysis: Option B reflects a mature horizon scanning response: capture the signal, assess it, integrate into the risk program and escalate with decision-useful options. A is reactive and ignores a credible signal. C is a disproportionate response before assessment. D misassigns a second-line activity to the third line and delays action. Correct answer: B.

11. Summary Points to Memorize

- Emerging AI risk is new, evolving or poorly understood, with scarce historical data and high uncertainty.
- Horizon scanning is a systematic, continuous, multi-source process to detect weak signals and feed strategic foresight into the AI risk program.
- Core steps: scope, sources, collect, analyze, assess and prioritize, respond and integrate, communicate, review.
- Key techniques: PESTLE, trend analysis, scenario planning, Delphi, risk radars, cross-impact analysis.
- Assessment dimensions: likelihood, impact, velocity, proximity, uncertainty, preparedness.
- Outputs must flow into the risk register, scenarios, KRIs, controls and board reporting.
- Governance: second line owns, first line contributes, third line assures, board sets appetite and demands foresight.
- On the exam, favor structured, forward-looking, integrated, cross-functional and appropriately frequent approaches, and always tie responses back to risk appetite and governance.