Tonight, we harden every door. Every night stands on its own.
How you do anything is how you do everything.

AI ExchangeBuilding Phoenix's AI community, together. Workuity hosts tonight, and we present Founder Weekend side by side.
A scripted AI bill. A forged webhook. A leaked key.
Every route: who, check, limit.
Row level security. Masked emails. Encrypted tokens.
The 21st search gets a 429
Forged webhooks get a 400
User A can't see user B
✓ 9 routes, 3 cost money, 1 has no auth check
claudeWho should be able to call /api/leads?
youAnyone, but insert only.
✓ Wrote SECURITY.md: 6 threats, 8 routes
Map every route, what it costs and who may call it, interview me, then write SECURITY.md. No secrets, no code changes. Wait for my OK.
✓ Better Auth limiter · database storage
✓ lib/ratelimit.ts · 20 searches an hour per user
200 ×20 · 429 ×5
claudeThe UI shows "Try again in 12 minutes."
Turn on Better Auth's limiter with database storage, wait while I set up Upstash, then limit the AI and signup routes per user and prove it with 25 calls.
better-auth.com/docs/concepts/rate-limit
Every rule starts as Log.
Each rule's conditions, action and undo.
In the dashboard, after 10 minutes on Log.
Exclude the cron routes from deny rules.
Write FIREWALL.md: an IP rate limit on the AI routes, a deny rule, an /admin allow list, cron excluded, all starting as Log. I publish them, then you update SECURITY.md.
vercel.com/docs/vercel-firewall/vercel-waf
unsigned request → 400
repeated svix-id → 200, skipped
✓ cron secret · timingSafeEqual
claudeThe approve route checks the plan owner before resumeHook.
Verify Resend webhooks on the raw body, reject stale requests, skip repeats with a 200, compare the cron secret in constant time, and check the approval route before resumeHook.
resend.com/docs/webhooks/verify-webhooks-requests
claudeThe app connects as the owner role. It bypasses RLS.
Proposed: app_user, 3 policies, FORCE on 3 tables
youRun it on the branch.
user A → user B's dinners: 0 rows
Propose a least-privilege app role, forced RLS policies and per-request set_config, wait while I create the role on a Neon branch, then prove user A can't see user B.
neon.com/docs/guides/rls-query-execution
⚠ lib/ai.ts:41 logs the request body
⚠ signed_up sends email to PostHog
✓ leads_masked · j***@gmail.com
claudeEvents now carry user ids only.
Find every PII leak by file and line, add a masked leads view, strip emails from PostHog with property_denylist and before_send, and log ids only.
posthog.com/docs/product-analytics/privacy
AES-256-GCM in lib/crypto.ts with a versioned key I generate myself, encrypted OAuth tokens, a proposed migration for old rows and a rotation plan.
✓ .env*.local ignored · never committed
✓ .env.tpl · 9 op:// references
⚠ RESEND_API_KEY is Config, not Secret
claudeRe-save it as a Secret, then rotate it.
Check .gitignore and git history, find hardcoded keys by location only, write a 1Password .env.tpl, and give me the Vercel and GitHub checklist.
1password.dev/cli
deepseek/deepseek-v4.1-flash · 14 files
HIGH · app/api/photo/route.ts:22 · no size limit
LOW · lib/email.ts:9 · logs recipient
claudeBoth are real. The size check goes before the AI call.
A script sends the branch diff and SECURITY.md to DeepSeek V4.1 Flash via AI Gateway for typed findings, then Claude checks each one against the code.
vercel.com/ai-gateway/models
Rules from SECURITY.md
in .cursor/BUGBOT.md
Cursor reviews each PR for bugs and security issues.
Cursor's Security Reviewer agents, set up in Automations.
You type it in Claude Code before every push.
Turn SECURITY.md into .cursor/BUGBOT.md, wait while I connect Bugbot, trigger it on a PR, then check every Done when item.
Also in the series: Part 10, Production Monitoring. Know the moment it breaks.
An Arizona AI hackathon. Teams of 1-4.
See the hackathon ↗