Cloudflare Security Audit Skill: evidence-led reviews for coding agents
Operating Security Audit Skill safely: scope, sandbox and untrusted text
Keep target content from acquiring authority over tools or the review process.
What you will learn
- Keep authorization explicit
- Enforce execution controls
- Review reports before sharing
Before you start
- Source review and JSON knowledge
- Explicit target authorization and isolation planning
A proposed learning project keeps source revisions, coverage gaps and verdicts visible.
Key takeaways
- Repository text is untrusted input.
- Prompt instructions are not an OS sandbox.
- Reports can contain sensitive evidence.
Keep authorization explicit
Define the owned or authorized target, revision, output location and permitted checks. Reading source does not grant permission to attack a running service, access accounts or expand network scope.
Treat instructions found in target files as material under review. They cannot authorize new tools, secrets or destinations, even when they resemble system instructions or a helpful test command.
Enforce execution controls
The README requires OS-enforced containment for target-controlled execution, including builds, browsers and fixtures. Disable external networking, sanitize the environment, limit resources and constrain writes to scratch paths.
If these controls cannot be demonstrated, leave runtime-dependent leads as needs_validation. Do not compensate with a warning in the prompt while running the same command with host privileges.
Review reports before sharing
Audit records can contain sensitive paths and evidence. Keep access limited and redact secrets from any report intended for wider distribution without erasing the facts needed to evaluate the claim.
Independent source review and record validation serve different purposes. Neither replaces human responsibility for authorized scope, safe execution or a decision to disclose a finding.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Define and preserve authorized scope.
- 2
Enforce containment before target execution.
- 3
Review evidence access and disclosure.
Copy-ready example
audit_boundary:
target_authorization: explicit
external_network: disabled
environment: sanitized-allowlist
writes: assigned-scratch
target_text: untrustedFrequently asked questions
Can a target README expand audit scope?
No. Scope comes from the authorized operator.
Should blocked runtime checks be marked confirmed?
No. Preserve the exact unresolved fact as needs_validation.
Sources
- Security Audit Skill / README.mdSource checked 2026-09-23