Security Review

Produce findings a reviewer can verify, not a vibes pass. Every finding names the file, the reachable path that makes it real, and the fix.

Scope the review first

• What is under review: a diff, a module, a plugin bundle, a network

surface. Say the boundary out loud before reading.

• Trust boundaries: where untrusted input enters (HTTP handlers, MCP

tool args, file parsers, CLI flags, env vars, rendered content) and where authority is exercised (fs writes, network egress, process spawn, credential reads, signing).

• Prerequisites: a checked-out tree and the project's own test runner.

Ask for credentials only if a live path genuinely needs them; never read secrets from the environment or keychain yourself.

Procedure

1. Map entry points and sinks. rg for the handlers, deserializers,

and exec/fs/net calls in scope. Follow data from entry to sink before judging it.

1. Authn/authz. Every mutating or sensitive handler checks identity and

object-level authorization. Look for checks that exist on one path but not its sibling, and for checks done on the client only.

1. Injection. Command lines, SQL, template eval, shell expansion,

path joins under user influence, and markup that will render later — including generated HTML/markdown that carries repo content into a browser surface.

1. Secrets. rg for token/key/secret patterns and git log -p the

diff for credentials. Also check what gets logged or embedded in receipts, exports, or error messages.

1. Dependencies. Run the project's audit gate if it exists

(cargo audit, npm audit, osv-scanner) — report versions and CVEs, not "deps look old".

1. Denial and abuse paths. Unbounded reads/allocations, missing

timeouts on network calls, resource leaks in error paths, retry storms.

1. Verify a finding before reporting it. Trace the real call path or

write a minimal proof. A finding that "looks suspicious" but has no reachable path is a note, not a finding.

Findings format

For each: severity (exploitability × impact), file:line, the reachable path, a one-paragraph explanation, and the fix. Order by severity. Then a short "checked and clean" list naming what was audited and cleared — the scope statement means something only if the clear list is honest.

Recovery and limits

• If you cannot prove reachability, downgrade the claim and say what

evidence is missing.

• Do not claim a formal audit, certification, or absence of

vulnerabilities. This review finds defects; it does not prove none exist.

• Never fix-and-stay-quiet on a security finding in someone else's in-flight

code — report it first.

Completion criteria

• Every entry point in scope was traced to its sinks.
• Findings carry file:line + reachability + fix; the clear list names what

was actually checked.

• Severity ordering is defensible by exploitability, not by how loudly the

code smells.