"AI audit" gets used as if it's one thing. It isn't. An audit of a single LLM's injection resistance and an audit of a multi-agent system's tool permissions are checking entirely different risks. Here's the actual taxonomy, the difference between an audit and ongoing monitoring, and five direct questions that tell you whether your systems are ready for one.
Scope note: for the step-by-step process and checklist of running an audit, see How to Audit an AI Model Before Deployment →. This guide is about which type of audit you actually need and whether you're ready for it.
An AI audit is a structured evaluation of an AI system's safety, security, reliability, fairness, compliance, and operational readiness. It looks at how a model was trained, how it behaves in production, how it handles adversarial input, how it interacts with data and APIs, and how it's monitored and controlled — distinct from ongoing monitoring, which observes behavior continuously rather than evaluating it at a fixed point in time.
Different systems need different audits. Picking the wrong type wastes effort checking risks that don't apply and misses the ones that do.
Prompt injection resistance, jailbreak vulnerability, unsafe output generation, and data leakage risk — the right scope for a standalone model or chatbot with no execution authority.
Autonomous decision-making, tool usage behavior, workflow execution safety, and API interaction control — the right scope once a system can take real actions, not just generate text.
Policy compliance, access control enforcement, human oversight mechanisms, and accountability structures — focused on whether the right controls exist, independent of whether they're being actively monitored.
Alignment with external frameworks like EU AI Act or NIST AI RMF, or with internal governance policy — the right scope when the driver is regulatory or contractual obligation.
Operational risk, cybersecurity exposure, adversarial attack surface, and systemic vulnerabilities — the broadest scope, typically run before a major deployment decision or after a near-miss incident.
| AI Audit | AI Monitoring | |
|---|---|---|
| Timing | Periodic evaluation | Continuous observation |
| Depth | Deep, structured assessment | Real-time signal tracking |
| Primary focus | Compliance-oriented | Operational-oriented |
| What it catches | Structural and design-time gaps | In-the-moment anomalies and drift |
Both are necessary for mature governance, not interchangeable. An audit alone misses everything that happens between audits; monitoring alone lacks the structured depth needed to catch design-level gaps. See AI Behavior Monitoring → for the continuous side of this pairing.
Not every AI system needs the same level of scrutiny. A system needs auditing when it makes decisions affecting users or business outcomes, accesses sensitive or regulated data, interacts with external systems or APIs, operates autonomously without human approval, or influences financial, legal, or operational workflows. The more of these apply simultaneously, the higher the audit priority and the broader the scope should be.
The HexTyx AI Security Assessment evaluates your system against all five audit types and tells you where the real gaps are.