The AI risk map, read by category

Most AI risk conversations are anecdotal: a single failure, a single headline, a single demo that goes wrong. What teams rarely have is a map — a structured way to see the full surface of what can go wrong before they ship. The AI risk map organizes dozens of failure modes into nine families that every AI team, founder and enterprise leader should be able to name.

The value is not the specific acronyms. It is the discipline of treating risk as a system with categories, instead of a list of incidents you react to one at a time. Below I walk through the nine families and the risks that live inside each.

How to read the map

The map groups individual failure modes into nine risk families. Three of them, model, data and security, are mostly technical. Three, governance, operational and business, are about how the capability is run and sustained. The last three, human and ethical, monitoring and control, and agentic, are where the newest and least understood risks concentrate. A mature program needs coverage across all nine, not depth only in the ones closest to engineering.

Model risks

These come from the model itself: how it learns, generalizes and behaves as inputs change.

• Hallucination (HAL): generates false or misleading answers. Ji et al., 2022
• Bias (BIA): produces unfair or one-sided outputs. Mehrabi et al., 2021
• Data drift (DRF): input data changes over time.
• Model drift (MD): performance slowly degrades.
• Overfitting (OVF): memorizes instead of generalizing. Zhang et al., 2017
• Underfitting (UND): fails to learn useful patterns.
• Error propagation (ERR): small errors spread across workflows.
• Uncertainty (UNC): confident but wrong answers. Guo et al., 2017
• Personal data exposure (PII): sensitive training data resurfaces. Carlini et al., 2021

Data risks

These come from the data pipeline that feeds and trains the model.

• Source risk (SRC): unverified sources reduce trust. Sambasivan et al., 2021
• Mislabeling (MIS): wrong labels damage learning. Northcutt et al., 2021
• Duplicate data (DUP): repetition reduces reliability. Lee et al., 2022
• Synthetic data risk (SYN): fake data creates weak behavior. Shumailov et al., 2023
• Consent risk (CON): data used without permission.
• Data leakage (DAT): training data leaks into responses.
• Data quality loss (DQL): poor data creates unreliable outputs.

Security risks

These are adversarial: someone is actively trying to misuse or break the system.

• Jailbreak risk (JBR): users bypass safety limits. Zou et al., 2023
• Prompt injection (PJI): malicious prompts manipulate behavior. Greshake et al., 2023
• Adversarial attack (ADV): inputs designed to fool the model. Goodfellow et al., 2015
• Data exfiltration (EXT): the system leaks confidential information.
• API abuse (API): attackers misuse connected services.
• Token theft (TOK): credentials or keys are exposed.
• Access risk (ACC): the wrong users reach restricted systems.
• Supply chain risk (SUP): third-party tools introduce vulnerabilities.

Governance and compliance risks

These appear when AI runs ahead of the rules, ownership and documentation around it.

• Governance failure (GOV): AI lacks proper oversight. NIST AI RMF, 2023
• Regulatory risk (REG): the system violates legal requirements. EU AI Act, 2024
• Compliance risk (CMP): AI breaks industry standards.
• Documentation gap (DOC): usage is poorly recorded.
• Ownership gap (OWN): no clear person owns AI outcomes.
• Policy failure (POL): internal rules are unclear.
• Audit gap (AUD): decisions lack proper records.
• Explainability gap (EXP): decisions are hard to understand. Rudin, 2019

Operational risks

These are about running the system reliably and affordably in production.

• Scaling risk (SCL): the system fails under higher demand.
• Cost overrun (CST): usage becomes too expensive.
• Latency risk (LAT): responses become too slow.
• Reliability risk (ROB): behavior is inconsistent in production.
• Deployment risk (DEP): poor rollout creates failures. Paleyes et al., 2022
• Integration risk (INT): AI breaks existing workflows. Sculley et al., 2015
• Rollback gap (ROL): teams cannot safely reverse failures.
• Monitoring gap (MON): issues go unnoticed. Sculley et al., 2015

Business and reputation risks

These hit value, trust and the brand rather than the code.

• Trust gap (TRU): people avoid using AI systems.
• Strategy risk (STR): adoption lacks clear direction.
• AI hype risk (HYP): expectations exceed real capability. Goldman Sachs, 2024
• Reputation risk (REP): mistakes damage the brand.
• Customer trust loss (CUS): users stop trusting outputs.
• ROI failure (ROI): investment brings weak returns. RAND, 2024
• Revenue risk (REV): failures hit business results.
• Vendor lock-in (VEN): the business depends too much on one provider.
• Job displacement fear (JOB): teams fear being replaced. Frey & Osborne, 2017

Human and ethical risks

These are about the effect on people and society.

• Fairness risk (FAI): AI treats groups unequally. Barocas et al., 2023
• Manipulation risk (MAN): AI influences users unfairly.
• Ethical risk (ETH): AI creates harmful social outcomes.
• Safety risk (SAF): AI causes real-world harm.
• Human-in-the-loop gap (HITL): critical decisions lack review. Parasuraman & Manzey, 2010

Monitoring and control risks

These are gaps in how the system is observed, tested and verified.

• Evaluation failure (EVAL): testing misses real-world issues. Liang et al., 2022
• Red teaming gap (RED): risks are not stress-tested. Ganguli et al., 2022
• Validation gap (VAL): outputs are not properly checked.
• Metric blindness (MET): teams track the wrong metrics.
• Alert failure (ALR): critical issues trigger no alerts.
• Logging gap (LOG): AI actions are not traceable.
• Automation bias (AUT): humans trust AI too quickly. Parasuraman & Manzey, 2010

Agentic AI risks

These are the newest family, appearing only once AI can act, not just answer.

• Kill switch gap (KILL): the AI cannot be stopped quickly.
• Feedback gap (FBK): user feedback never improves the system.
• Agent autonomy risk (AGT): agents act beyond intended limits. Chan et al., 2023
• Tool misuse (TLS): agents use tools incorrectly. Barrett et al., 2023
• Unauthorized action (ACT): an agent performs restricted operations.
• Memory risk (MEM): agents store sensitive context.
• Multi-agent failure (CHAIN): one agent's error affects others.
• Loop failure (LOOP): agents repeat actions endlessly.
• Goal misalignment (GOAL): the agent pursues the wrong objective. Amodei et al., 2016
• Delegation risk (DEL): agents pass tasks without control.

The real strategic question

The wrong question is: "Which of these risks should worry us most?" Picking a favorite risk is how programs end up well-defended on one side and wide open on another.

The better question is: "Which of these nine families do we currently have no owner, no metric and no control for?" Most teams discover that their coverage is uneven: strong on model and security, thin on governance, monitoring and, increasingly, agentic risk. The map is useful precisely because it makes those blind spots visible.

A pragmatic direction

You do not mitigate seventy risks at once. You make the surface legible, then close the most exposed gaps first.

That usually means:

• map your real use cases onto the nine families and mark where you have no owner
• treat monitoring, evaluation and governance as first-class concerns, not afterthoughts
• harden security and model behavior with red-teaming and evaluation loops
• put explicit human-in-the-loop controls where impact or uncertainty is high
• before deploying agents, confirm you can stop, observe and constrain them

A risk map does not remove risk. It turns a vague anxiety about "what could go wrong with AI" into a finite, ownable list, and that is the first step from reacting to incidents toward designing for resilience.