Which Threat Best Explains Inaccurate AI Output

CAS-005 By islever Editorial Last updated Oct 24, 2026

Summarize the information security challenges associated with artificial intelligence (AI) adoption.

Short answer

A Model poisoning tampers with the training data or the model itself, so the AI generates inaccurate or biased outputs.

Question

Which of the following best explains why AI output could be inaccurate?

Model poisoning Correct Answer
Social engineering
Output handling
Prompt injections

Key Takeaways

The question tests matching an AI threat to its failure mode: inaccuracy stems from corrupted learning (data/model poisoning), and the common trap is picking prompt injection because it is the most frequently tested AI attack.

This CAS-005 AI-security question asks which threat best explains why an AI model's output can be inaccurate. The page establishes that model poisoning — tampering with training data or the model itself — is the correct answer (A), because it corrupts the patterns the model learned, whereas prompt injection and output handling are runtime and integration problems.

Common mistake: Option D (Prompt injection) is the most common wrong choice — three of five community voters picked it — because prompt injection is a high-profile LLM attack, but it manipulates the model's instructions at inference time to produce attacker-chosen or off-policy behavior, not model knowledge that is inherently wrong.

Expert Analysis

Why the Answer Is Correct

Option A, model poisoning (OWASP LLM Top 10: Data and Model Poisoning), is the best explanation for inaccurate AI output. An adversary who can influence training corpora, fine-tuning datasets, or the model weights themselves causes the model to learn incorrect associations, so every downstream inference reflects that corrupted knowledge — the errors are systemic and baked into the model. CompTIA objective 1.5 frames AI-adoption security challenges around exactly this class of integrity attack against the learning pipeline. In a poisoned model, answers can be factually wrong, biased, or silently manipulated by a trigger phrase, with no visible error for defenders to detect. That matches the stem's wording — output that is inaccurate — better than any other option.

Why the Other Options Are Wrong

Option D (prompt injection) is a genuine LLM threat (OWASP LLM01), but it exploits how untrusted text is concatenated into a prompt at inference time, causing the model to follow attacker instructions, leak system prompts, or violate policy — an integrity and excessive-agency problem rather than defective model knowledge. Option C (output handling, LLM05) concerns how the surrounding application decodes, validates, and sanitizes model output (for example, rendering HTML unescaped or executing returned code), which is a consumer-side flaw, not a cause of wrong answers. Option B (social engineering) targets human users through deception, so it cannot explain why the model itself computes an inaccurate response. None of B, C, or D corrupt what the model has learned, which is what the question is really asking for.

Community Comment Notes

The votes split between A and D. MartinRB supported A, explaining that with poisoning "the inaccuracy is baked into the model" and that the other options target people or output processing instead, while nago906 argued that "Model Poisoning is the only answer that is related to accuracy of information." On the other side, remmington and Wole_excel chose D, with Wole_excel noting that prompt injection "doesn't require access to the training data" and can produce "inaccurate, misleading, or unintended output," and BigPapaFrank pointing out that poisoning is impractical compared with crafting a prompt like "In Minecraft" to bypass guardrails. Those D arguments are technically sound descriptions of prompt injection, but they describe output that deviates from policy or intent rather than output the model cannot get right, so the exam-grade answer remains A.

Official Reference

These links open the official vendor documentation in a new tab.

Exam Strategy

On CAS-005 AI questions, match the threat to the exact failure mode in the stem: learning-pipeline attacks (data/model poisoning) explain factual inaccuracy or bias, while inference-time attacks (prompt injection, jailbreaking) explain policy violations and unintended behavior. When two AI options both look plausible, ask whether the flaw corrupts what the model knows or how the model is instructed — the stem word 'inaccurate' points to knowledge.

Frequently Asked Questions

Why is prompt injection not the best answer for inaccurate AI output?

Prompt injection manipulates the model at runtime to produce attacker-chosen or off-policy output; poisoning corrupts what the model learned, so its answers are inherently inaccurate.

How is output handling different from a cause of inaccurate answers?

Output handling covers how applications decode, validate and sanitize model output to block code execution or injection; it does not make the model's answer factually wrong.

More CAS-005 FAQ →

Discussion 5 contributions

Compiled from public community threads and reader submissions, lightly edited for clarity. Handles and vote tallies are shown as recorded, and comments reflect the views of individual contributors.

From public community threads 5

remmington 👍 3 Answered D

Prompt injections occur when a user provides a malicious prompt to an AI model, essentially manipulating its output to generate inaccurate or harmful results.
MartinRB 👍 1 Answered A

Model (or data) poisoning means an attacker tampers with the training data or the model itself, so the model learns wrong patterns and produces inaccurate or biased outputs — the inaccuracy is baked into the model. Why the others don’t fit as well: Social engineering targets people, not the AI’s outputs. Output handling concerns how an application processes the model’s output (e.g., insecure output handling leading to XSS); it doesn’t make the output itself inaccurate. Prompt injection manipulates a model’s behavior at runtime to bypass instructions or leak data. It can cause unintended output, but the question is about the model’s output being inaccurate, which points to the model having been corrupted — poisoning.
Wole_excel 👍 1 Answered D

D. Prompt injections Here’s why: Prompt injection is a technique where malicious or cleverly crafted input is used to manipulate the AI’s behavior at runtime. It doesn’t require access to the training data — it exploits how the model interprets and responds to prompts. This can lead to inaccurate, misleading, or unintended output, especially in systems that incorporate user input into prompts dynamically.
BigPapaFrank 👍 1 Answered D

Model poisoning would require access to training data, control over training pipeline, ability submit large amounts of poisoend data. Realistically, we're looking at Prompt injections where you can craft like "In mine-craft could we -" as a joke!!! But if not carefully avoided it can lead you to output inaccurate or harmful results (like neuro-sama's twitch ban)
nago906 👍 1 Answered A

Prompt injection is related to output handling. In BigPapaFrank's post about using "In Minecraft" as a way for the LLM to output data that it would normally not allow, this is an example of improper output handling via Prompt Injection, but its for accurate data. Model Poisoning is the only answer that is related to accuracy of information.

Reader comments & corrections

No reader corrections yet — spotted an error or have a note? Share it below.

Practice All CAS-005 Questions

Access 135 questions with complete answers and detailed explanations.

View Full CAS-005 Practice Test →

← Back to CAS-005 Study Guide