ShipSafe: the pentest-and-fix loop for vibe-coded apps before they go live
The problem
AI coding tools optimize for a working demo, not for authentication correctness or database access control, and the apps they generate go straight to production with real users. The documented failures are now a pattern, not anecdotes: 170 Lovable apps exposed user data through broken Supabase row-level security (CVE-2025-48757), Moltbook leaked 1.5 million API tokens and 35,000 emails in January 2026 from AI-generated code with no RLS, and a scan of 5,600 vibe-coded apps found 2,000+ vulnerabilities and 400+ exposed secrets. The builders have no security team, and the scanners built for security teams do not speak their workflow.
The idea
Why now
sapper-v2 (30 Oct) packages an autonomous pentest legion with a commander and worker swarm; OpenAI's GPT-6 Cyber (5 Nov) is a hacking model built for vetted defenders; Cloudflare's security-audit-skill (June) turns a coding agent into a security auditor; and Lovable itself now runs an automatic security scan on publish — the platform vendors have validated the need, but their scanners are locked to their own platform and stop at reporting.
What it combines
sapper-v2's autonomous pentest legion + GPT-6 Cyber's defender-side hacking model + Cloudflare's security-audit-skill + sandbox-mcp's fresh Windows Sandbox machines for safe testing. A pentest legion without a fix loop is a report nobody reads, a hacking model without a sandbox is unsafe to run, and an audit skill without re-testing cannot confirm the fix — together they are a loop that attacks, fixes, and proves the fix, on every deploy.
MVP
Weekend MVP: a CLI that takes a deployed vibe-coded app URL, runs five fixed attacks (unauthenticated table read, secret in JS bundle, auth bypass on one endpoint, IDOR on one resource, open registration), and emits a pass/fail report with the exact request that succeeded. Skip: auto-fix PRs, multi-app support, and the public badge.
How it makes money
The moat
Signature-based scanner incumbents (Intruder, Nessus) cannot bolt an autonomous attack-and-fix loop onto a scanner without rebuilding around agent exploitation and code-writing — their architecture finds known patterns, it does not attack, fix, and re-verify. The fix corpus is also a data advantage: every closed hole teaches the legion the next variant, and a new entrant starts with no exploit history.
Distribution
B2B2C through the vibe-coding platforms themselves — a pre-publish check inside Lovable, Bolt, Replit, and Base44, plus a GitHub Action for everyone else. The wedge is founders about to put real users on an AI-built app, reached where they already deploy.
Why it wins
Intruder, Detectify, and Escape scan and report — they do not write the fix or re-attack to confirm it, and they price per target for security teams. Lovable's own scanner only works inside Lovable. ShipSafe is the only loop that closes the hole in the builder's own repo and proves it closed, for the indie founder who has no security team at all.
Risks
Biggest risk: the legion's attacks against a live app could corrupt real user data or get the user banned by their host. The MVP de-risks it by attacking only a staging clone by default and requiring an explicit flag — with a written warning — for production URLs.
Build it with
- sapper-v2: autonomous pentest legion with an AI commander and worker swarmThe autonomous pentest legion — commander plus worker swarm that attacks like a real adversary.
- OpenAI GPT-6 Cyber: a hacking model for vetted defenders onlyThe defender-side hacking model that powers the exploitation step.
- Cloudflare security-audit-skill: turn your coding agent into a security auditorTurns a coding agent into a security auditor — the fix-writing half of the loop.
- sandbox-mcp: fresh Windows Sandbox machines for testing apps, via MCPFresh isolated machines so attacks run safely away from production data.
Repo to start from
shipsafe — a pre-publish pentest loop for vibe-coded apps: attack, auto-fix, re-attack, and issue a pass badge.
Evidence
- Vibe-coded app security in 2026: 170 exposed Lovable apps and 1.5M keys
- Vibe coding platform Zeabur shutting down after two back-to-back breaches
- Intruder pricing: Essential from $149/mo; AI pentests from $3,500/test
- Vibe coding security: SusVibes benchmark, 57% work but only 11.8% secure
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.