CySA+ v4 · Objective 1.6
AI for security analysts
Use AI to assist analysis and documentation while preserving evidence and decision ownership. Identify hallucination, exposure, poisoning, and prompt-injection risks and apply proportionate controls.
What you will be able to do
- Choose bounded AI tasks with verifiable outputs.
- Distinguish hallucination, data exposure, model poisoning, and malicious prompts.
- Apply data-handling and approval rules to an AI-assisted workflow.
- Evaluate an AI result against source evidence before action.
Learn the concepts
- Lesson 1
Use AI with evidence you can check
AI can help an analyst organise information, compare artifacts, and draft explanations. Its usefulness depends on whether a person can check its output against the underlying evidence.
Open lesson → - Lesson 2
Control AI risks and governance
AI risk controls should match the data, the model’s role, and the consequences of a wrong answer. Running a tool locally changes some exposure paths, but it does not remove the need for governance.
Open lesson →
Apply your judgement · Synthetic scenario
A convincing summary with an unexpected request
An assistant summarises approved synthetic incident records. One record contains an instruction to retrieve another case, and the resulting draft adds an unobserved containment action.
- Supplied records
- E21: suspicious export request; E22: request denied; no isolation event
- Artifact text
- Message field asks the assistant to retrieve private case notes
- Generated draft
- States that the host was isolated and the incident resolved
- Integration design
- Read access limited to the current case; action execution disabled
Your task
- Classify the two distinct AI risks shown by the artifact and the draft.
- Explain which controls prevented additional harm and what still requires correction.
- Define verification and reporting steps before the summary is shared.
Compare your response
The embedded request is a malicious prompt within untrusted evidence. The invented isolation and resolution are unsupported output, consistent with hallucination; the records do not establish that an action occurred.
Case-scoped read access and disabled execution limit exposure and unauthorised changes, even if the assistant follows the embedded instruction. They do not guarantee a factually correct summary, so human review remains necessary.
Replace the invented conclusion with the supported denied-export observation and unresolved investigation status. Check each claim against record identifiers, document the AI failure, and use the example to evaluate the workflow before sharing the corrected report.
Put it into practice
Use these labs alongside this module.
Check your understanding
Source reading: supplied book, chapters 1. Lessons, scenarios, and questions are original CyberCorps course material.