Academy learning group 4 · Academy module (not an ISO clause) M08 · Chapter 8 of 8
Investigate AI-specific security signals
About 5 minutes · ISO/IEC 27090:2026 · Published AI-security guidance
Session only. Enable saving or download a backup from this course’s hub.
Completion is self-reported and does not certify understanding.
See it · Try it · Recall it
Picture the topic
Remember this: Observe the behavior, preserve the version, then test the hypothesis.
Explore the anchors
Collect relevant protected observations.
Concept matching activity
Match this task or description to its concept:
Correlate data model and service evidence.
Choose a concept before checking the explanation.
Detection laboratory
Positive means the detector raised an alert. True/false compares that decision with known reality.
False negative — An attack exists. No alert was raised.
| Detector | Attack present | No attack |
|---|---|---|
| Alert | True positive | False positive |
| No alert | False negative | True negative |
Change the threshold
Eight fictional events have fixed known truth. An alert fires when the synthetic score is at least the threshold. The score is not a probability of attack.
TP 2 · FP 1 · FN 2 · TN 3. Precision 67%. Recall 50%.
Precision = TP ÷ (TP + FP). Recall = TP ÷ (TP + FN). Lower thresholds can increase detection and false alarms; this tiny dataset does not establish real-world performance.
A quick way to study this topic
- Hide the explanation and redraw these three anchors from memory.
- Explain the difference between the anchors in your own words.
- Apply the distinction to the worked example below. Check the full lesson for assumptions and exceptions.
- Recall it again tomorrow, then after a few days. A memory hint is a retrieval cue, not a replacement for understanding.
Original academy illustrations and fictional exercises. These are learning simulations, not a real virtual machine, network assessment or measure of job readiness. Activities reset when you leave or reload; they do not change quiz scores.
What this chapter explains
- Concepts and distinctions
- Apply the process and collect evidence
Concepts and distinctions
AI-security observations can include unusual inference patterns, unexpected output changes, model-store access, data-pipeline modifications and tool actions outside normal use. A changed output may reflect a malicious input, an authorized update, random variation or a data-quality problem. Preserve relevant context before deciding which hypothesis the evidence supports.
Monitoring needs defined use cases, known limitations and an operating response route. A detector’s false positives and misses affect workload and risk. Do not interpret confidence scores as proof of a named attacker or use a few synthetic events to claim measured production effectiveness. Logs also need proportionate retention, access control and privacy protection.
Apply the process and collect evidence
Record the model and configuration version, relevant data lineage, timestamps and affected service state. Use approved containment options that consider business and safety consequences. Coordinate model, data, operations, security and privacy owners. Validate a trusted recovery state and test whether the identified failure path is addressed before reopening the same workflow.
| Concept | What to remember |
|---|---|
| Telemetry | Collect relevant protected observations |
| Investigation | Correlate data model and service evidence |
| Response | Contain safely and validate recovery |
Worked example
An assistant begins exposing unexpected document fragments after a model-service update. The first responder wants to delete all stored prompts and immediately retrain the model. Preserve authorized evidence and relevant versions, restrict the exposed path under response authority and investigate competing explanations. Deleting records or retraining before investigation can destroy useful evidence and introduce another unreviewed change. Apply retention and privacy requirements while preserving necessary facts.
Check your understanding
What should responders preserve when investigating unexpected AI behavior?
Show explanation
Relevant versions configuration lineage and authorized event evidence. These facts help distinguish attack hypotheses from updates and data-quality changes.
Source and study boundary
Official standard overview. The reference identifies the topic being taught. Explanations and examples are original academy material; a topic link does not establish independent approval or sufficient exam mastery.