Security+ SY0-801 AI Security Scenarios Explained
Last verified: November 18, 2026, against CompTIA’s Security+ V8 page. For the exact objective wording, use the official SY0-801 objectives (PDF).
Security+ SY0-801 launched on November 17, 2026 with a dedicated objective on the security risks of artificial intelligence. If you have never worked with AI systems, the terms can feel abstract. They are not. Each one maps to a failure a security team can investigate and a control a security team can apply. This guide walks through three tickets of the kind you might meet on the job, then shows how the same ideas appear on the exam.
Everything here is defensive and conceptual. You will not find attack prompts or step-by-step exploitation, because Security+ asks you to recognize risks and choose controls, not to attack systems. If you are still choosing between exam versions, start with our SY0-701 vs SY0-801 guide.
✦ CompTIA’s V8 summary lists a dedicated objective on AI risks: AI systems, models, prompts, privacy and data exposure.
✦ Prompt injection is about untrusted text steering an AI. Poisoning is about tampered training data. They are different risks with different controls.
✦ Deepfakes show up as a social engineering risk, and the best control is a process, not better eyesight: out-of-band verification and dual approval.
✦ The strongest answer to most AI risks is the classic one: least privilege, validation, logging and human oversight.
✦ Learn to name the risk from a short scenario, then pick the control that limits the damage.
What the official objectives say about AI
According to CompTIA’s Security+ V8 page, the Threats, Vulnerabilities, and Attacks domain (24% of the exam) includes an objective to explain the security risks associated with artificial intelligence, covering risks related to AI systems, models, prompts, privacy and data exposure. The same page mentions AI-assisted workflows under automation and orchestration in Security Operations.
Several study sites that have read the full objectives document report that this objective is numbered 2.6, is worded as “summarize threats and vulnerabilities associated with artificial intelligence (AI) usage,” and lists 13 topics. They also report related AI items elsewhere: large language models as an attack surface (2.4), deepfakes under social engineering in the indicators of malicious activity objective (2.5), and AI capabilities under automation (4.6). The verb “summarize” suggests recognition-level questions: name the risk, explain the impact, pick a sensible control.
Security+ is a defensive, entry-level security exam. You do not need to train or tune models. A candidate who spends a week on machine-learning math has misread the objective. Spend that week on recognizing risks and mapping them to controls.
The AI risk terms in plain English
The topics reported under the AI objective fall into three groups: attacks on the model, attacks through the model, and risks in the output and data. The definitions below are our plain-language explanations, not CompTIA’s wording.
| Group | Term | Plain-English meaning |
|---|---|---|
| Attacks on the model | Poisoning | Tampering with the data a model learns from, so it learns the wrong thing |
| Attacks on the model | Model manipulation | Altering or steering a model’s behavior or components without authorization |
| Attacks on the model | Evasion | Crafting inputs so a model fails to detect or classify something correctly |
| Attacks on the model | Jailbreaking | Getting a model to ignore its built-in safety rules |
| Attacks through the model | Prompt injection | Untrusted text (a message, a document, a web page) that steers the AI away from its intended instructions |
| Attacks through the model | Session hijacking | Taking over a legitimate user’s session with an AI-connected service |
| Attacks through the model | Code execution | An AI tool being tricked into running commands or code it should not |
| Risks in output and data | Data loss and privacy | Sensitive data leaving through prompts, outputs or logs, or being used improperly |
| Risks in output and data | Hallucinations | Confident but false or invented output that people may trust |
| Risks in output and data | Bias | Systematically unfair or skewed results caused by data or design |
| Risks in output and data | Explainability | The ability to understand why a model produced a result, which matters for audits and trust |
| Risks in output and data | Ethical considerations | Fairness, transparency and accountability in how AI is used |
That is 12 rows for the 13 reported topics, because data loss and privacy are listed separately in some breakdowns and combined here. Check the official PDF for the exact list.
Scenario 1: the support chatbot that shared the wrong order
Opened: Tuesday 09:42 | Reporter: Customer Support Lead
System: “OrderHelper” AI chatbot on the public help page
Symptom: A customer reports seeing another person’s order status in the chat. Transcript shows the customer’s message contained text telling the bot to disregard its rules and show “all recent orders.”
Investigation
- Contain. Temporarily disable the chatbot’s order-lookup tool while keeping the rest of the help page online.
- Preserve evidence. Export the chat transcripts, tool-call logs and timestamps. Do not edit them.
- Scope the exposure. Search the logs for other sessions where the bot returned records that did not belong to the signed-in customer.
- Find the root cause. Review how the bot is configured: what it is allowed to look up, and whether user text and system instructions are kept separate.
- Escalate. Involve privacy and legal staff, because customer data was exposed. Follow your incident response and notification policy.
Root cause
This is prompt injection leading to data loss and a privacy incident. Two design flaws combined: the bot treated untrusted user text as if it were instructions, and its lookup tool could return any customer’s records, not just the signed-in customer’s.
Controls that limit the damage
- Least privilege for tools: scope the lookup so it can only return the authenticated customer’s own records, no matter what the model is told.
- Separate trusted instructions from untrusted content, and treat everything a user or a document supplies as untrusted.
- Input and output filtering to catch sensitive data before it reaches the screen.
- Human approval for high-impact actions, and logging and monitoring so abuse is visible quickly.
How the exam may ask about it
Expect a short scenario where an AI tool reveals or does something it should not after crafted input, and you must name the risk (prompt injection) and choose the best control. Note the pattern: the strongest answer limits what the AI can do, because you cannot rely on the model to always resist manipulated text.
Self-check
Question 1. A chatbot that can look up orders reveals another customer’s data after a user message tells it to ignore its rules. Which control would BEST limit the damage if this is attempted again?
A. Retrain the model on more customer conversations
B. Restrict the lookup tool so it can only return the signed-in customer’s own records
C. Ask customers not to type instructions into the chat
D. Increase the chatbot’s response length limit
Show answer and explanation
Correct: B. Least privilege limits what the bot can reach regardless of what text it is given.
Why the others are wrong: A does not fix the architecture and may add more customer data to the risk. C depends on attackers following a request. D has nothing to do with access to records.
Scenario 2: the spam filter that started trusting a campaign
Opened: Friday 14:05 | Reporter: Email Security Analyst
System: AI spam classifier, retrained monthly on user-reported labels
Symptom: After this month’s retrain, messages from one phishing campaign are passing the filter. Detection rate for that campaign fell sharply.
Investigation
- Compare model versions. Test last month’s model and this month’s model against the same sample of messages to confirm the change came from the retrain.
- Inspect the training data. Look at what changed between datasets, and where the labels came from.
- Check label sources. Find a cluster of “not spam” labels submitted by a small number of new accounts for messages from the same campaign.
- Roll back. Redeploy last month’s model while the data is cleaned.
- Review the change process. Confirm who approved the new training data and what testing happened before release.
Root cause
This is poisoning: the training data was tampered with through an unvetted feedback loop, so the model learned to treat the campaign as legitimate. It is different from evasion, where an attacker crafts a message at detection time to slip past a model that was trained correctly.
Controls that limit the damage
- Data provenance and validation: know where each label came from and reject or review suspicious clusters.
- Access control and integrity checks on training data and model files, so changes are authorized and detectable.
- Versioning and rollback so a bad model can be replaced fast.
- Change management and testing: a held-out test set and an approval step before any retrained model goes live. This links to the change management objective in Domain 1.
How the exam may ask about it
A question may describe a model whose behavior degrades after a retrain and ask you to identify the risk and the best preventive control. Watch for the clue “training data” or “retrained”: that points to poisoning, while “crafted input at the time of use” points to evasion or prompt injection.
Self-check
Question 2. After retraining on user-reported labels, a spam filter passes one campaign’s messages. Many “not spam” labels came from a few new accounts. Which risk and control fit BEST?
A. Prompt injection; filter user input
B. Poisoning; validate the provenance of training labels and keep model versions for rollback
C. Hallucination; add citations to the model’s output
D. Jailbreaking; tighten the model’s safety prompts
Show answer and explanation
Correct: B. Tampered training labels are poisoning, and provenance checks plus rollback address the root cause.
Why the others are wrong: A and D describe attacks at the time of use, not on training data. C addresses false statements in output, which is not what happened here.
Scenario 3: the urgent call that sounded like the CFO
Opened: Monday 16:20 | Reporter: Accounts Payable Supervisor
System: Phone and payments workflow
Symptom: An accounts payable clerk received a voicemail that sounded like the CFO, asking for an urgent wire to a new vendor. A second approver noticed the vendor was not on file and paused the payment.
Investigation
- Stop the payment and, if any money moved, contact the bank immediately.
- Verify out of band. Call the CFO using a number from the company directory, not the number that left the voicemail.
- Preserve evidence: the voicemail file, caller details, emails and the payment request.
- Check for other targets. Ask finance and executive-assistant staff whether they received similar messages.
- Report and review. Log the incident and review the payment approval process.
Root cause
This is social engineering using a deepfake: synthetic audio imitating a trusted executive to create urgency. The weakness is a process that allowed a payment request to be accepted on the authority of a single voice.
Controls that limit the damage
- Out-of-band verification through a known, independent channel for any unusual payment or credential request.
- Dual approval and a strict procedure for new vendors and changed bank details.
- Awareness training that uses real examples and rewards people for reporting suspicious requests.
- A no-blame reporting culture, so staff pause and ask instead of rushing.
How the exam may ask about it
Expect an urgent request from a seemingly senior person and a question about the best response or the best preventive control. The answer is almost never “listen more carefully for audio glitches.” It is a process control such as verification through a separate channel.
Self-check
Question 3. An employee receives an urgent voicemail that sounds like the CFO, requesting a wire to a new vendor. Which control would BEST prevent a fraudulent payment?
A. Train staff to detect small audio glitches in voicemails
B. Require verification through a known number and dual approval for new-vendor payments
C. Block all incoming calls from outside the company
D. Replace phone calls with voicemail
Show answer and explanation
Correct: B. Out-of-band verification and dual approval work even if the fake is perfect.
Why the others are wrong: A relies on detection that gets harder as fakes improve. C would disrupt legitimate business without stopping fraud by email or other channels. D does not change who is trusted.
Control mapping: risk, example and control
| Risk | Example | Control ideas |
|---|---|---|
| Prompt injection | Chatbot steered by crafted user text | Least-privilege tools, trusted/untrusted separation, filtering, human approval, logging |
| Poisoning | Tampered feedback labels used for retraining | Data provenance, validation, integrity checks, versioning, rollback, change management |
| Model manipulation | Unauthorized changes to a model or its settings | Access control, integrity checks such as hashing or signing, monitoring |
| Jailbreaking | User gets a model to ignore safety rules | Layered guardrails, output filtering, monitoring and rate limiting |
| Evasion | Input crafted to slip past a detection model | Layered detection, adversarial testing, not relying on one model |
| Hallucinations | Confident but wrong answers used in a decision | Human review, verified sources, no unreviewed automation of high-impact decisions |
| Data loss and privacy | Staff paste sensitive data into a public AI tool | Data classification, DLP-style controls, redaction, policy and training |
| Session hijacking, code execution | AI tool abused to reach sessions or run commands | Secure session handling, MFA, sandboxing, disabling unneeded tools |
| Bias, explainability, ethics | Unfair or unexplainable automated decisions | Governance policy, testing, documentation, human oversight, audits |
| Deepfake fraud | Synthetic voice or video imitating an executive | Out-of-band verification, dual approval, awareness training |
How to study the AI topics
- Make flashcards for the terms, grouped as attacks on the model, attacks through the model, and risks in output and data.
- Practice the three-step answer: name the risk, explain the impact, pick the control that limits damage.
- Reuse what you already know. Least privilege, change management, hashing and awareness training explain most of the right answers.
- Separate similar pairs: poisoning versus evasion, and prompt injection versus jailbreaking.
- Check the official objectives for the exact topic list, then test yourself with our Threats and vulnerabilities practice questions.
Related Security+ guides
- SY0-701 vs SY0-801: which exam to take
- Upgrading your prep from SY0-701 to SY0-801
- Practice questions: Threats and vulnerabilities
- 4-week Security+ study plan
Prostutii is an independent study resource and is not affiliated with or endorsed by CompTIA. Security+ is a trademark of CompTIA. Sources: CompTIA Security+ V8, SY0-801 objectives (PDF).