AI risks in finance

Most lists of AI risks are either a glossary with no governance in them, or a controls framework with no plain explanation of the risk. This taxonomy sits in between. It names the risks a finance firm actually has to manage, explains each one in plain terms, and points to how the MAS AI Risk Management Guidelines and the MindForge toolkit expect you to handle it.

Each risk below links to its own page: what it is, why it matters in finance, the controls that catch it, and the questions to ask. I keep this openly available, last reviewed 24 October 2026; if you want to add or correct something, drop me a note at gary@quaintitative.com. For the full step-by-step treatment, see the guide for risk and control functions and the complete guide.

How generative models fail

The failure modes that come with open-ended, generated output. These are the newest for most firms, and the ones a predictive-model framework does not cover.

• Hallucination risk. The model produces fluent, confident output that is not supported by any source. The same style carries a correct figure and an invented one.
• Prompt injection risk. Instructions hidden in the content a model reads steer it into doing something it should not. The attack arrives through the input, not the code.
• Data leakage risk. Sensitive data leaves the firm through a prompt, a third-party model or a log, or surfaces in an output it should not.

How models degrade

Risks that show up over time, after a system is live, rather than at the point it ships.

• Model drift. The world moves away from the conditions the model was built for, and performance decays without warning until it matters.
• Reproducibility and traceability risk. You cannot reconstruct how a given output was produced, so you cannot investigate it, defend it or fix it.

Fairness and human factors

Risks that sit between the system and the people using or affected by it.

• Bias and fairness. The system treats similar people differently in ways that are unfair or unlawful, often through the data it learned from.
• Over-reliance and automation bias. People defer to the system because it is confident and convenient, and stop applying the judgement the control assumed they would.

Security, opacity and concentration

Risks that touch the firm's security, its ability to explain itself, and its dependence on a small number of providers.

• AI security risks. New attack surfaces specific to AI, from poisoned training data to model theft and adversarial inputs.
• Explainability and opacity risk. The firm cannot explain, to a customer or a regulator, why the system produced a given outcome.
• Third-party model concentration risk. Much of the industry depends on the same few foundation models, so one provider's failure or change propagates widely.

Already covered in depth

Some of these risks already have fuller treatments on this site, so the taxonomy points to them rather than repeating them.

• Agentic AI risks. What changes when a system takes actions and chains steps on its own, and how to govern the path rather than a single output.
• Third-party AI. Holding bought AI to the same standard as built AI, and what an AI Card does and does not tell you.
• Making AI safe in practice. The controls that actually reduce generative-AI risk, and how you tell they are working.
• Where AI risk hides from the second line. The places risk accumulates out of sight of oversight.