Security Sweep

Point it at a SaaS app and it finds the holes that cost money or leak tenants, not a generic lint pass. It concentrates fire on the handful of classes with real blast radius (someone pays $0, someone reads another tenant's data, a key leaks), proves each one is actually exploitable before it says a word, ranks by damage, and hands back a fix per finding. Low-risk fixes land under a green ratchet; anything touching a live payment or auth path is queued for the user, never auto-edited.

Core Insight: A security finding is worthless until it's proven reachable. Scanners and smart models both over-report: they flag the shape of a bug without tracing whether an unauthenticated request can actually get there. The hard part isn't listing OWASP categories; it's building the exploit path in your head (who, unauthenticated or cross-tenant, hits which route, bypassing which check) and rejecting the plausible-but-wrong noise. Report only what you'd bet on. And because the highest-value classes (billing and auth) are also the ones where a wrong "fix" causes an outage or a revenue leak, those changes are proposed, not applied: findings auto, payment/auth-critical edits to the decision inbox.

Preconditions / guards

1. Know the stack before you grep: Next.js route handlers vs. Edge middleware, Supabase RLS vs. service-role keys, Stripe vs. RevenueCat webhooks each fail differently. Read how auth, billing, and the data layer are wired first; a finding written against the wrong mental model is a phantom.
2. Green ratchet on any applied fix: A fix that reds a test reverts. Never weaken a test to pass.
3. Payment-flow and auth-critical changes are ask-user: Produce the finding and the recommended fix, but write the change to <project>/.autopilot/DECISIONS.md (recommended option first) and DO NOT auto-edit a live billing or session path. A bad auto-fix here is an outage or a free-money bug.
4. Read-only against real services: Never fire test webhooks at prod, never exercise a live charge path to "confirm". Reason the exploit; reproduce only in a test-mode/local sandbox.

The audit classes (where the fire goes)

Full checklist with Next.js/Supabase/Stripe-specific failure modes in references/CLASSES.md. The seven:

1. Payment / billing bypass: entitlement granted without a settled charge; client-trusted price/quantity/plan; webhook-before-payment ordering; refund/ downgrade that never revokes access; RevenueCat receipt trusted unsigned.
2. Webhook integrity: missing/incorrect signature verification (Stripe constructEvent, RevenueCat auth header), no replay/idempotency guard, raw-body mismatch from a framework that already parsed the body.
3. Auth gaps: authn vs authz confusion, missing server-side check behind a hidden UI control, JWT not verified / alg:none / wrong audience, session fixation, cookies missing HttpOnly/Secure/SameSite, getSession() trusted where getUser()/getClaims() is required.
4. RLS / multi-tenant isolation: RLS disabled or USING (true), a service-role key reachable from a user-facing path (bypasses RLS entirely), tenant_id taken from the request instead of the verified session, a query that forgets the tenant filter.
5. Secrets exposure: keys in source / committed .env / logs, a server secret prefixed NEXT_PUBLIC_ and shipped in the client bundle, service-role key in browser code.
6. IDOR / broken access control: object id from the URL with no ownership check, sequential ids, mass-assignment letting a request set role/tenant_id/is_admin.
7. Input → injection surface (secondary): only where it crosses into the above (SQL via raw query bypassing RLS, SSRF reaching internal billing APIs).

The Pipeline (staged: issue each phase only when the prior finishes)

Phase 0  MAP        Read how auth, billing, the data layer, and webhooks are wired
                    (guard #1). List every trust boundary: each route handler, each
                    webhook endpoint, each Supabase query path, each entitlement
                    gate. This map is the attack surface.
Phase 1  SWEEP      Walk the seven classes against the map. For each candidate,
                    write the EXPLOIT PATH: who (unauth / other tenant / lower role)
                    hits what, bypassing which absent check, to what effect. No
                    exploit path = not a finding yet.
Phase 2  VERIFY     Adversarially re-check EVERY candidate against the real code:
                    open the cited file/line, confirm the missing check is actually
                    missing (not enforced one layer up in middleware / RLS / a
                    guard). Default to "not a finding" if you can't trace the path.
                    Doubtful + high-stakes → multi-model-triangulation if installed,
                    else a second independent trace. Phantoms die
                    HERE, before anything is written or changed.
Phase 3  RANK       Order survivors by blast radius × reachability: unauth remote
                    money/tenant-leak first, then authed cross-tenant, then
                    same-tenant escalation, then info-leak. Severity is damage ×
                    ease, not category name.
Phase 4  REMEDIATE  Split by risk:
                    • Low-risk (add missing authz check on a non-billing route, set
                      a cookie flag, move a secret, add an ownership filter) → fix
                      under the green ratchet, each with a regression test that fails
                      on the vulnerable code and passes on the fix.
                    • Payment-flow / auth-critical / RLS-policy change → DO NOT EDIT.
                      Write finding + recommended fix to .autopilot/DECISIONS.md.
Phase 5  PROVE      For each applied fix, prove the exploit is now blocked (test-mode
                    repro / the regression test), not just that the suite is green.
                    Re-sweep the touched surface; a fix often exposes an adjacent gap.

Output (what orchestrators get back)

• Ranked verified findings: each with: class, severity (blast radius × reachability), the exploit path, cited file:line, and the concrete fix.
• Applied fixes: low-risk only, green-ratcheted, each with its regression test.
• <project>/.autopilot/DECISIONS.md: appended: every payment-flow / auth- critical / RLS-policy change, recommended option first, for the user to approve. Same inbox project-autopilot and idea-forge drain: one inbox, not three.

Contract (what orchestrators rely on)

• bug-zero routes its security class (auth bypass, webhook integrity, RLS, secrets) here in the SWEEP phase; this skill returns verified findings into bug-zero's own verify/fix loop.
• project-autopilot routes "Billing/payment/webhook security" here. Findings produced AUTO, payment-flow CHANGES ask-user via the decision inbox.
• ship-it calls it as a pre-ship hardening gate.
• Most specific skill wins: chosen over generic audit-lens for any billing / auth / webhook / multi-tenant code.

Skills it composes with

These are optional; use any that are installed, otherwise the provider's own docs cover the same ground. Verify doubtful findings with multi-model-triangulation; for billing flows pair with saas-billing-patterns-for-stripe-and-paypal and stripe-checkout to know the correct shape before calling a deviation a bug; supabase for RLS/auth specifics; prove fixes with testing-real-service-e2e-no-mocks (Stripe test mode, real-DB tenant-isolation tests). It adds the security lens + exploit-path verification; these supply the domain detail.

THE EXACT PROMPT

/security-sweep
Target: <project path or subsystem (e.g. the billing + webhook routes)>
Scope:  <whole app | billing/webhooks | auth/session | tenant isolation>
Stack:  <e.g. Next.js + Supabase RLS + Stripe webhooks>   (optional; it will infer)

Anti-Patterns

• Reporting a finding with no exploit path. "Looks risky" is not a vulnerability; trace who reaches it bypassing which check, or drop it.
• Calling a missing check a bug without confirming it isn't enforced one layer up (middleware, RLS policy, a shared guard). The classic phantom.
• Auto-editing a live payment or session path. That's an outage / revenue-leak risk; it goes to the decision inbox, full stop.
• Firing real webhooks or live charges at prod to "confirm". Repro in test mode only.
• Ranking by OWASP category name instead of blast radius × reachability. An unauth remote tenant-leak outranks a same-tenant self-XSS every time.
• Banking a fix with no regression test that fails on the vulnerable code.
• Trusting getSession() / client-supplied price / client-supplied tenant_id and not flagging it because "the UI never sends that". The attacker isn't using your UI.
• One generic scan instead of walking the seven classes against the actual trust map.

References

• CLASSES.md: the seven audit classes expanded, each with Next.js / Supabase / Stripe / RevenueCat failure modes, what to grep for, and the exploit question that decides whether it's a real finding.