security-audit

name	security-audit
description	Run a security audit on the current branch's changes. Use when the user asks to audit, security-review, or check for vulnerabilities in pending changes. Spawns the security-auditor agent against the diff.
user-invocable	true

/security-audit — Audit pending changes for security issues

Arguments: $ARGUMENTS (optional: a specific file or package path to focus on)

---

What to do

You are acting as the security-auditor agent. Review the pending changes (or the specified path) for security issues specific to grew's threat model.

1. Get the diff

git diff main...HEAD --stat
git diff main...HEAD

If $ARGUMENTS names a file or package, also read it in full:

cat $ARGUMENTS

If there are no changes vs main, audit the working tree instead:

git diff HEAD

2. Identify the trust surface

For each changed file, classify it:

• Network boundary: pkg/downloader, pkg/tap, pkg/osvdev, any HTTP client usage
• Filesystem boundary: pkg/safepath, pkg/fsutil, pkg/cellar, pkg/caskroom, pkg/linker, pkg/relocation, archive extraction
• Process boundary: pkg/sandbox, pkg/sudo, any exec.Command usage
• Cryptographic: pkg/signing, pkg/snapshot, hash verification, key loading
• CLI input: cmd/*/, flag parsing, argument validation

3. Apply grew-specific checks

For each trust boundary touched, check:

Downloads & verification

• Both SHA256 and SHA512 are checked (not just one)
• Hashes verified before extraction, not after
• HTTPS enforced; HTTP URLs rejected at parse time
• HOMEGREW_ALLOWED_HOSTS allowlist respected

Path safety

• All paths through pkg/safepath before use
• Archive extraction rejects ../ and absolute paths (Zip Slip)
• Symlink creation validates target is within prefix

External commands

• -- separator passed to git, hdiutil, tar, launchctl
• No exec.Command("sh", "-c", <user-input>)
• Arguments passed positionally, not via shell string interpolation

Sandboxing

• Build scripts and post-install hooks run under Seatbelt
• Sandbox profile does not allow writes to $HOME

Privileges

• sudo only for initial prefix setup
• --unsafe flag gated by devmode build constraint, not env var

Self-update

• Binary patches dual-hash verified before apply
• OSV check fails closed (error = abort)
• Atomic replacement: no partial binary left on failure

4. Check for OWASP-class issues

Even outside grew-specific checks, flag:

• Command injection via unsanitized user input in subprocess calls
• Path traversal in any file operation
• SSRF via user-controlled URLs
• Privilege escalation via file permission mistakes

5. Output findings

For each finding:

Severity: Critical / High / Medium / Low / Info File:line: exact location What: one sentence Why it matters: concrete exploit or impact Fix: specific code change

End with: N critical, N high, N medium, N low findings.

If no issues found, say so explicitly.