CARBON — Engineering ConfidenceUser Guide
CHAPTER 22 / 60 Pro

/carbon-security-review

Review AI systems for prompt injection, untrusted channels, data leakage, cross-tenant access, tool permissioning, unsafe side effects, containment, privacy, physical security, and user trust.

Diego · AI chatbotSecurity & resilience

When to use it

AI inputs, retrieval or tool calls introduce new trust boundaries.

Before you start

Map untrusted text to privileged actions and data access. Use isolated adversarial fixtures, check containment, and keep external side effects outside the test unless explicitly approved.

Try this in your coding agent.

/carbon-security-review Review the support chatbot for prompt injection and unintended data access.

Replace sample paths with your project. Slash names are shorthand; your host may use a plugin prefix or a skill picker. Invocation help →

ILLUSTRATIVE WALKTHROUGH

Support-agent trust boundaries

An expanded teaching example, not a recorded or live execution. The report stays visible while the replay explains how to read it.

  1. 01Read context
  2. 02Investigate
  3. 03Show evidence
  4. 04Report

Coding-agent conversation

Read context

I’m reviewing support-agent trust boundaries. I’ll keep the scope and permissions explicit.

Investigate

Follow the weakest evidence link and record what remains blocked.

Show evidence

Here is the evidence trail to inspect: Untrusted document → instruction-like content → proposed account-change tool call.

Report

Retrieved text can influence a privileged tool request. Use a sandbox and approved test scope before executing adversarial cases.

CARBON

Example report

Illustration
Answer vs supporting source
Answer vs supporting source — TestBucks teaching screen

TestBucks course screen with illustrative analysis. The screenshot supplies context; the demo does not establish a newly observed bug.

  1. Question / intent
  2. Reference evidence
  3. Scoring & uncertainty

Illustrated outcome

Retrieved text can influence a privileged tool request.

Evidence to inspect

Untrusted document → instruction-like content → proposed account-change tool call.

Still untested / not established

Use a sandbox and approved test scope before executing adversarial cases.

Ready. Read the full example or play its steps.

What a useful result contains

A bounded claim

For support-agent trust boundaries, state the target, revision, environment and actual work performed. Keep planned, executed, blocked and untested work separate.

An inspectable trail

Untrusted document → instruction-like content → proposed account-change tool call. In a real run, retain the source, trace, screenshot or fixture behind the conclusion.

The qualification

Use a sandbox and approved test scope before executing adversarial cases.

The next decision

Decide whether the evidence calls for a focused check, a reviewed fix or more context. Do not convert a plausible example into a verified finding.

Adapt the example to your project

  1. Replace support-agent trust boundaries with a real feature, path or decision in your project.
  2. State the expected rule and its source. Include safe fixtures, constraints and a budget where appropriate.
  3. Ask for exact evidence and remaining gaps. Approve consequential actions separately.
/jay Review the /carbon-security-review result. Separate observed evidence, assumptions and untested scope, then recommend the next useful check.

Where to go next

/carbon-eval →/carbon-incident →

Content checked against the local command definition: skills/carbon-security-review/SKILL.md. Examples adapted from the homepage’s TestBucks walkthroughs. Availability is edition- and host-dependent.