Assess an AI feature's exposure on paper

applied · 75 min · Objective 4.6

Task

Assess the attack surface an LLM-backed feature adds to an application — on paper, by reasoning about it, not by attacking a live model — because CompTIA scopes objective 4.6 as explaining prompt injection and model manipulation, and the safe, accurate exercise is an exposure assessment rather than a jailbreak.

Steps

  1. Describe a realistic feature on paper: an application that passes user-supplied text to an LLM and acts on the result — a support assistant, a document summariser, a code helper.
  2. In /tmp/assessment.md, map where untrusted input reaches the model: directly from the user, and indirectly through content the model is asked to process (a document, a web page, an email).
  3. Reason about direct prompt injection: what a user could try to make the feature do outside its intent, and what the blast radius is if it works.
  4. Reason about indirect prompt injection: how instructions hidden in processed content could reach the model, which is the harder case.
  5. Reason about model manipulation in outline — poisoning, extraction, evasion — and where each would apply to this feature.
  6. Write the exposure conclusion: what the feature can reach if the model is steered, and what controls reduce it.

Verify

grep -cE "^##\s+(Direct injection|Indirect injection|Model manipulation|Exposure)" /tmp/assessment.md
grep -ciE "untrusted input|reaches the model|blast radius" /tmp/assessment.md
grep -ciE "on paper|no model|not attack|assess" /tmp/assessment.md

The first count must be at least 4 — the assessment covers the required areas. The second must be non-zero: you identified where untrusted input reaches the model and what it can reach, which is the heart of the assessment. The third must be non-zero: the assessment is explicit that it reasons about exposure rather than attacking a model, which is both the safe reading and the accurate one.

Notes

CompTIA's own verb for this objective is explaining, not performing — the only explanatory verb among six attack objectives — so the accurate exercise is assessing an AI feature's exposure, not walking through a jailbreak. Where an LLM-backed feature widens an application's attack surface is a real, report-worthy question; answering it on paper is both safer and truer to how the exam frames it. No model was attacked in this lab.

This is an independent study companion for CompTIA PenTest+ PT0-003 and is not produced by or endorsed by CompTIA.