security

name	security
description	Rails/Grape security review: authz, injection, output sanitization, SSRF. Triggers: "is this vulnerable", "SQL injection", "XSS", "Brakeman warning". Do NOT use for: secret scans.
user-invocable	false
effort	medium

Security

Iron Laws

1. Authorization must be explicit at every boundary.
2. Never interpolate untrusted input into SQL.
3. Never mark untrusted content as HTML safe.
4. Secrets come from environment or secure credentials, not source files.
5. Security-sensitive Sidekiq work must honor transaction boundaries.

Secret Detection

Use /rb:secrets to scan for leaked credentials with betterleaks:

• /rb:secrets — Scan current directory
• /rb:secrets --git — Scan git history
• /rb:secrets --validate — Validate secrets against live APIs

Install betterleaks: brew install betterleaks

See references/betterleaks-integration.md for detailed configuration.

Evidence Mode

Every finding MUST carry evidence_mode:

Mode	Use when
static-signal	Grep / pattern match. Lowest trust.
runtime-confirmed	Reproduced via existing test OR read-only Tidewave introspection.
configuration-risk	Config-only issue (yml, env, initializer).
requires-human-validation	Threat-model / business-context input needed.

Prefer runtime-confirmed. Refuse to emit findings without evidence_mode.

runtime-confirmed is read-only

NEVER synthesize destructive code (DELETE/UPDATE/DROP/TRUNCATE/rm/mv, new tests that mutate shared state, live network calls to non-mock endpoints, config edits) to prove a finding. Allowed: existing tests, read-only Tidewave queries, log inspection. If non-destructive reproduction impossible → downgrade to static-signal or requires-human-validation.

Defense in depth: CC sandbox network denylist

For infra-layer egress restriction (cloud metadata endpoints, known exfiltration targets), use CC 2.1.113+ sandbox.network.deniedDomains. Runs beneath plugin permission rules; applies to all Bash tool calls. See /rb:permissions → "Network egress restriction" for the settings.json shape.

Gotchas

• Brakeman false-positive flood. Default Brakeman config flags many false positives. Triage by evidence_mode: static-signal Brakeman hits are LOWER priority than runtime-confirmed SQL injection in failing test.
• evidence_mode missing. Every finding MUST carry evidence_mode enum (defined in § Evidence Mode above). Refuse to emit findings without it.
• Pundit policy-scope omission. Authorize EVERY controller action (Iron Law 13). before_action :authorize alone is insufficient — verify per-action policy_scope or per-action authorize call.
• html_safe / raw on user input. Iron Law 14 violation. Search \.html_safe|raw\( in changed files; flag any with non-trusted source.

References

Need	Reference
betterleaks CLI integration + git-history scan	${CLAUDE_SKILL_DIR}/references/betterleaks-integration.md
Pundit/CanCanCan policies, password hashing, session auth	${CLAUDE_SKILL_DIR}/references/authentication.md
explicit authorization patterns + multi-tenant scopes	${CLAUDE_SKILL_DIR}/references/authorization.md
input validation, file-upload magic bytes, path traversal, XSS sanitization, SQL safety, command-injection prevention	${CLAUDE_SKILL_DIR}/references/input-validation.md
OAuth identity resolution + token refresh + multi-provider	${CLAUDE_SKILL_DIR}/references/oauth-linking.md
Rack::Attack rate limiting + composite keys + Redis limiter	${CLAUDE_SKILL_DIR}/references/rate-limiting.md
CSP, CSRF, HSTS, security headers, Brakeman + bundle-audit CI wiring	${CLAUDE_SKILL_DIR}/references/security-headers.md
SSRF, secrets management, supply chain, CORS, safe deserialization, file-upload content-type	${CLAUDE_SKILL_DIR}/references/advanced-patterns.md