shir-bruchim - Code review — Review code for security vulnerabilities
- Setup hooks — Configure safety hooks to block dangerous operations
- File protections — Set up protected-path blocking
- Deep review — Comprehensive security audit across 10 dimensions with checklists
Wait for response.
Scan for OWASP Top 10 against this canonical checklist; for each match capture file:line + severity + remediation per the per-finding format below.
- SQL injection: parameterized queries everywhere (never f-string SQL)
- XSS: HTML-encode user output; Content-Security-Policy header
- CSRF: tokens on state-changing requests (or SameSite cookies)
- Auth bypass: authorization checked on every protected route (IDOR)
- Sensitive data: passwords hashed (bcrypt/argon2/scrypt), secrets in env vars
- Error messages: no stack traces or internal details in production
- Dependencies: no critical CVEs in direct or transitive deps
For each finding:
- File path and line number
- Severity: Critical / High / Medium / Low
- Explanation of the vulnerability
- Specific remediation code
Never mark complete without:
- Authentication/authorization flows checked
- All user inputs validated at system boundaries
- Sensitive data (passwords, tokens, PII) handled correctly
- No hardcoded credentials or secrets
- SQL queries parameterized (no string concatenation)
block-dangerous-commands.sh — Primary hook. Blocks rm -rf /, dd to disks, fork bombs, mkfs, and git force pushes. Exits 2 (hard stop). block-dangerous-bash.sh — Compatibility alias. Delegates to block-dangerous-commands.sh; use when your hooks.json references the "bash" filename. protect-files.sh — Blocks writes to ~/.ssh, ~/.gnupg, ~/.aws/credentials, and any paths in PROTECTED_PATHS or ~/.claude/protected-paths.txt. Exits 2 (hard stop).
To install (Claude Code) — add to ~/.claude/settings.json:
{
"hooks": {
"PreToolUse": [
{
"matcher": "Bash",
"hooks": [{"type": "command", "command": "bash ~/.claude/hooks/block-dangerous-commands.sh"}]
},
{
"matcher": "Write|Edit",
"hooks": [{"type": "command", "command": "bash ~/.claude/hooks/protect-files.sh"}]
}
]
}
}
Copy hook files: cp hooks/block-dangerous-commands.sh hooks/block-dangerous-bash.sh hooks/protect-files.sh ~/.claude/hooks/ && chmod +x ~/.claude/hooks/*.sh
Instead, add a "never do" list to the project's AGENTS.md (or GEMINI.md) and to the AI's always-on rules:
NEVER run: rm -rf /, git push --force, dd to /dev/sd*, mkfs, fork bombs.
NEVER write to: ~/.ssh/, ~/.gnupg/, ~/.aws/credentials, .env files with real secrets.
The safety is enforced by instruction, not by a hard block. For high-risk environments, prefer Claude Code with hooks.