- In short
- A system vulnerability walkthrough examines every entry point along the request and data path - user input, retrieved content, tool outputs, the model's own output, and the logs - asking at each what an adversary could do and which control currently stands in the way. Any point with no assigned control is a plausible attack surface regardless of how unlikely it seems, and the walkthrough pairs naturally with the guarded-path checkpoints already designed. Logs are themselves a data-path point that can leak sensitive fields even when the model output is clean.
Walk the path, not just the front door
Naming risk categories tells you what to look for; the system vulnerability walkthrough tells you where to look. The CCAR-P exam treats it as an analyse-level skill because the method is what turns a list of categories into a concrete assessment of one system. The technique is disciplined and simple: walk the request and data paths together, stop at every entry point, and ask two questions - what could an adversary do here, and which control currently stands in the way.
The reason to walk the whole path is that attacks do not politely confine themselves to the front door. Every place data enters or leaves the system is a candidate surface: user input, retrieved content, tool outputs, the model's own output, and the logs. A review that inspects only the initial request examines one of five points and declares the system assessed, which is exactly the incomplete review that passes and then fails in production.
- The system vulnerability walkthrough
- A structured review that walks every entry point on the request and data path - user input, retrieved content, tool outputs, the model's output, and the logs - asking at each what an adversary could do and which control stands in the way. Any point with no assigned control is a plausible attack surface, and the logs are a data-path point in their own right.
Every entry point gets the same two questions
The power of the walkthrough is its uniformity. At each point, you ask what an adversary could do and which control answers it. At user input, the adversary crafts a direct injection or an oversized payload; input screening stands in the way. At retrieved content, the adversary plants an indirect injection; retrieved-content screening stands in the way. At tool outputs, a tool response carries an instruction the model trusts; the same screening applies. At the model's output, the model produces something unsafe; output screening stands in the way. And at the logs, a sensitive field is written where it should not be; log minimisation and access controls stand in the way.
The rule that makes the walkthrough rigorous is this: any point with no assigned control is a plausible attack surface, regardless of how unlikely it seems. You do not get to skip a point because an attack there feels improbable - improbable is not the same as impossible, and the walkthrough's job is to make every surface explicit so the improbable ones are decided deliberately rather than ignored by default.
The logs are a surface, and they are the one people skip
The single most commonly omitted point is the logging pipeline. Reviews naturally gravitate toward the model - its inputs and its outputs feel like where the action is - and stop at "the model's response." But logs are a data-path point that can leak sensitive fields even when the model's output is entirely clean. A response can pass every content filter and still have written a Social Security number or a full account record into a log store along the way, where a different set of eyes and a different retention policy govern it.
Because data exposure through logging is independent of the model's behaviour, no output filter would ever catch it. That independence is precisely why the walkthrough forces you to treat the log as its own entry point with its own control. Skipping it is not a small omission - it is skipping the surface where a large share of real-world data-exposure incidents actually happen.
What the CCAR-P exam trips candidates on
The first trap is stopping the walkthrough at "the model's response" and never considering the logging pipeline as an exposure point. A scenario details careful input and output screening and asks whether the system's data is protected; candidates who accept the model-centred boundary miss the log. The credited reading names the logs as an unexamined surface where sensitive fields can leak with clean output.
The second trap is assuming a control that covers one entry point automatically covers a different one. A scenario shows strong input screening and asks whether retrieved content is protected; the credited reading is no - input screening does not inspect retrieved content, so that point still needs its own control. The walkthrough exists precisely to prevent one control's coverage from being silently extended to a point it never touches.
Worked example
You are asked to review a customer-support agent that retrieves answers from a partner knowledge base, issues refunds through a tool, and writes all activity to a request log. The team says the design is covered because they have input screening and an output filter. Walk the system and identify what their review missed.
Walk the five points. User input: covered by their input screening - direct injection and oversized payloads have a control. Model output: covered by their output filter - an unsafe response has a control. So far their two controls account for two points.
Now the three they missed. Retrieved content from the knowledge base is an entry point with no control: an indirect injection planted in an article reaches the model, and input screening never inspects retrieved content, so this surface is open. Tool outputs are similarly unscreened. And the refund tool itself is a side-effecting action with no authorization gate - tool and action abuse has no control standing in the way. Finally, the request log is a data-path point: it can capture sensitive fields even though the output filter passed clean text, so data exposure through logging is an open surface too.
Their mistake was two of the traps at once. They stopped at the model's response and never treated the log as an exposure point, and they let input screening's coverage stand in implicitly for retrieved content and tool outputs it never touches. The walkthrough surfaces all of it: retrieved-content and tool-output screening, deterministic tool-call authorization before refunds, and log minimisation with access and retention controls. Each open point becomes a row in the written risk assessment that follows.
Common misreadings to avoid
Misconception
A vulnerability review is complete once you have examined the model's inputs and its response.
What's actually true
Misconception
A control on one entry point, like input screening, also covers the other points.
What's actually true
How this shows up on the exam
Domain 5 items hand you an architecture and ask you to identify its vulnerabilities or what a review missed. The reliable method is to walk all five entry points - input, retrieved content, tool outputs, model output, logs - and, at each, name the adversary action and the control. Flag every point without a control, and never let one control's coverage extend implicitly to a point it does not touch. The logs are the point candidates forget.
This method operationalises the LLM system risk taxonomy by mapping each category onto a point on the path, and it pairs directly with the three-point guarded request path whose checkpoints are the controls you check for. Its output feeds the risk assessment as a written deliverable, where each open surface becomes a documented row with a mitigation and an owner.
A team says their agent is fully assessed because it has input screening and an output filter. The agent also retrieves from a knowledge base, issues refunds via a tool, and logs all activity. What did their review most likely miss?
People also ask
How do you walk an LLM system for vulnerabilities?
Which entry points does a vulnerability review cover?
Why include the logging pipeline in a vulnerability walkthrough?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
No videos curated for this concept yet
We are still curating the best official and community videos for this topic.
Official prep for this domain
Anthropic's own free prep module for this part of the syllabus, on the official prep course. Free with an Anthropic Academy sign-in.
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.