Chapters0/54 ▾
- 1.Architecture & Log Sources0/5
- 2.Indicators of Malicious Activity0/5
- 3.Investigation Tools & Threat Hunting0/5
- 4.Process Improvement, Automation & AI0/5
- 5.Vulnerability Scanning Methods0/5
- 6.Analyzing Scan Output0/5
- 7.Prioritization, Mitigation & Controls0/5
- 8.Attack Frameworks & the Response Process0/5
- 9.Incident Response Techniques0/5
- 10.Vulnerability Reporting & Communication0/5
- 11.Incident Reporting & Metrics0/4
Chapter 4 · Lesson 5 of 5
AI governance: accountability, audit trails and validation
1 min read
Governance questions ask who owns a decision and whether it can be explained later. A model's score is one input to a disposition, but the disposition itself must belong to a named analyst, because a wrongly closed alert has to be explained and a model cannot be held accountable.
The productive design lets the model rank and group alerts it judges benign while a person records the disposition that closes each group. Letting the model auto-close and reviewing closures at month end, or only in low-risk categories, still leaves cases closed with no human judgment behind them.
Auditability means the decision trail can be rebuilt months later: what was submitted to the tool, what it returned, which analyst accepted the conclusion and which evidence that analyst checked. Saving only the tool version does not help, because a model's answer is not reproducible the way a replayed query is. The same rule applies to AI-drafted detection rules.
Example: a model writes a rule that parses cleanly; before production, run it over 30 days of stored telemetry to see how often it matches and confirm that every field it references is actually collected. Letting live production alerts reveal the false-positive rate spends analyst attention to learn what a historical run would have shown.
Check yourself
1. Months after an incident, a regulator asks how the team concluded that no personal data left the organization. The case record says the decision followed an AI-assisted review, and nothing further about it was kept. What must the governance requirement be?
2. A platform feature can close any alert the model scores benign above a set confidence, with no analyst opening it. The operations manager must decide how to deploy the feature. Which arrangement is correct, and why?
Facts checked 2026-11-15 · Sources: CompTIA CySA+ (CS0-004) certification page and exam objectives, NIST AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1)
CompTIA CySA+ cheat sheet: every key fact on one page →