# bleek — Find every leak in your vibe-coded app (Full Documentation) > A specialized security scanner for applications built with AI code-generation tools, focused on the failure modes those tools introduce by default. ## About bleek bleek is a URL-based security scanner built specifically for applications created with vibe-coding tools — Lovable, Bolt.new, Cursor, Claude Code, v0, Replit, Windsurf, Base44, and similar. Founded by Ali Saeed under the Alenor brand, bleek identifies the specific vulnerability patterns these tools introduce when they generate code without explicit security prompts. The product name reads phonetically as "leak" — and that is the product. Every check looks for something that should not be exposed. ## The Problem We Solve Multiple independent security audits in 2025–2026 established the scale of the problem: - SupaExplorer scanned 20,052 vibe-coded URLs; 11% (2,217 domains) leak Supabase keys. - Beesoul / Retool: approximately 70% of Lovable apps ship with Row-Level Security disabled. - Escape.tech: 5,600+ vibe-coded apps analyzed, 2,000+ critical vulnerabilities, 400+ exposed secrets, 175 personal-data exposures. - Tenzai Security: every one of 5 AI tools tested introduced SSRF vulnerabilities in identical app builds; 69 total vulns. - Veracode 2025 GenAI Code Security Report: 45% of AI-generated code chose insecure implementations; 86% XSS failure rate; 70%+ Java security task failure rate. - CodeRabbit (470 GitHub PRs): AI-co-authored code introduces XSS at 2.74x the human-only rate. Documented breaches include: - Moltbook (Jan 2026): Fully vibe-coded social network. RLS not enabled. Public anon key in the bundle became an effective admin key. Within 72 hours of launch, 1.5 million authentication tokens and 35,000 user email addresses were exposed. - Lovable Exam App: 18,697 users exposed (4,538 university students) via inverted access-control logic — authenticated users blocked, unauthenticated users granted access. 170 Lovable apps shared this pattern (CVE-2025-48757). - Base44: Platform-wide authentication bypass — registration and OTP endpoints required no authentication; attackers could access any app via publicly-visible app_id. - Orchids: Zero-click remote code execution on a BBC journalist's machine. Company admitted missing 12 warning messages. - Enrichlead: Founder publicly stated app was built with "zero handwritten code" via Cursor AI. Users bypassed subscription paywalls, ran up API keys, random database entries appeared. Root cause: authorization checked client-side only. - Replit / SaaStr (Jul 2025): AI agent deleted 1,206 executive records and 1,196 company records during an ALL-CAPS code freeze issued by founder Jason Lemkin. ## Supported Platforms bleek works with applications built using: ### AI Code Generation Tools - Lovable (formerly GPT Engineer) — full-stack web apps - Bolt.new — StackBlitz-based AI coding - Cursor — AI-powered code editor - Claude Code — Anthropic's coding assistant - v0.dev — Vercel's AI UI generator - Replit — cloud IDE with AI assistance - Windsurf — Codeium's AI IDE - Base44 — vibe-coding platform - Antigravity, Augment Code, Devin, Cline, Cody, Tabnine, Amazon Q, Gemini Code, OpenAI Codex, GitHub Copilot ### Backend / Database Platforms - Supabase — PostgreSQL with Row-Level Security - Firebase — Google's app platform with Security Rules - PlanetScale — serverless MySQL - Neon — serverless PostgreSQL - Turso — embedded SQLite - Convex — backend platform - Appwrite — open-source backend - Airtable — spreadsheet database - MongoDB Atlas — managed MongoDB - Upstash — serverless Redis ### Hosting Platforms - Vercel - Netlify - Railway - Render - Fly.io - Cloudflare Workers ## What bleek Checks (Phase 1 — 14 checks) ### Passive URL-Only Checks 1. **Secrets in JS bundles** — Regex + signature matching for: Supabase service_role JWT, Supabase anon JWT (informational), OpenAI sk- and project- patterns, Anthropic sk-ant-, Stripe sk_live, AWS AKIA, GCP AIzaSy, GitHub ghp_, generic JWT, generic API tokens. Implementation: gitleaks + trufflehog patterns applied to all loaded JS/HTML/CSS files. 2. **Security headers** — Presence and correctness of: Content-Security-Policy, Strict-Transport-Security, X-Frame-Options or frame-ancestors CSP directive, X-Content-Type-Options, Referrer-Policy, Permissions-Policy. Implementation: HTTP HEAD/GET inspection + Mozilla Observatory v2 API for grade. 3. **Cookie flags** — Every Set-Cookie response header inspected for Secure, HttpOnly, SameSite. Implementation: HTTP response header parsing. 4. **TLS / cipher suite** — Protocol versions accepted (TLS 1.0/1.1 = fail), cipher suite quality, certificate validity. Implementation: sslyze CLI subprocess. 5. **CORS misconfiguration** — Detection of wildcard Access-Control-Allow-Origin combined with Access-Control-Allow-Credentials. Implementation: OPTIONS preflight inspection. 6. **Vibe-coding-platform fingerprinting** — Detect which AI tool built the app (Lovable, Bolt, v0, Base44, Cursor) via response headers, bundle paths, meta tags, framework signatures. Drives downstream platform-specific checks. ### Active URL Read-Only Checks 7. **Supabase RLS probing** — The flagship check (CVE-2025-48757). Pipeline: - Extract Supabase anon key from JS bundle (JWT regex). - Extract Supabase project URL (regex *.supabase.co). - Enumerate likely table names (users, profiles, accounts, orders, items, posts, messages, payments). - Probe `GET /rest/v1/?select=*&limit=1` with the anon key. - Observe: 200 with rows = RLS off or policy broken (CRITICAL). 200 with [] = RLS enabled (PASS). 401/403 = correct (PASS). 8. **Firebase test-mode rule probing** — Extract Firebase config (apiKey regex), then probe `GET https://.firebaseio.com/.json?shallow=true`. 200 with data = test-mode rules. 401 = correct. 9. **Exposed-file probes** — Nuclei template execution for: .env, .env.production, .env.local, .git/config, /backup.sql, /dump.sql, source-map files (*.js.map), /admin, /panel, /debug, /api-docs, /swagger.json, /.well-known/security.txt, /robots.txt leaked paths, /sitemap.xml leaked paths. 10. **Path traversal probes** — `../`, `%2e%2e%2f`, double-encoded variants against file-serving endpoints. 11. **Reflected XSS + SQLi probes** — Canary string injection on discovered query parameters. SQLi: single-quote then look for SQL error in response. Tool: OWASP ZAP baseline scan mode. 12. **Unauthenticated state-changing endpoint detection** — Enumerate routes from JS bundle, observe whether POST/PUT/DELETE methods return 2xx without an authentication header. 13. **Dependency vulnerability scan** — Retire.js on detected JS libraries + OSV.dev advisory lookup. 14. **Open-redirect + CWE-288 auth-bypass-via-alternate-path** — `?redirect=`, `?next=`, `?url=` parameter testing. Also: try `.json` variants of admin URLs, X-Forwarded-Host header tricks, direct REST that bypasses UI middleware. ## What bleek Outputs Each finding includes: - Severity: Critical, High, Medium, Low, or Info - Category: OWASP Top 10:2025 tag, OWASP API Top 10:2023 tag, OWASP LLM Top 10:2025 tag where applicable - CWE: corresponding Common Weakness Enumeration ID - Evidence: the exact response signature that triggered the finding (URL, HTTP status, sample response) - AI-Ready Fix: a structured Markdown block containing: - The literal config / code change for your framework (Next.js, Vercel, Cloudflare, Supabase RLS SQL, Firebase rules) - "Before applying" pre-flight checks - "Do NOT apply if" refusal conditions - "Safe-to-apply" flag for side-effect-free fixes - Designed for direct paste into Claude Code, Cursor, Windsurf, Copilot, or Lovable. ## What bleek Does NOT Do bleek is read-only by design. We never run destructive payloads or write to your data. We probe; we observe; we report. bleek cannot reliably detect the following from URL-only scanning (these require authenticated probing or repo access — see Human Checker tier): - IDOR / BOLA confirmation (OWASP WSTG explicitly requires multi-account testing) - Business-logic flaws (coupon stacking, negative-quantity carts, step-skipping, race conditions) - Payment-webhook signature verification correctness - Stored XSS / second-order injection (payload fires only after persisted data is viewed) - Authorization correctness at the data-property level (BOPLA) - LLM06 Excessive Agency blast-radius for agentic tools - Full system-prompt review (LLM07) - Compliance posture (GDPR / PCI / HIPAA) These eight categories are the defensible scope of the Human Checker tier. ## Tiers - **Free Scan** — $0 forever — Passive URL checks only. No signup. Email optional for the report. - **Deep Scan** — $19 one-time — All 14 Phase-1 checks. Supabase RLS probing, Firebase rule probing, injection probes, AI-ready fix instructions per finding. - **Human Checker** — book a call — For the 8 categories above. A named human reviews and ships fixes as pull requests you approve. - **Continuous Protection** — coming soon — Daily re-scans, diff alerts, trust badge auto-revoke on new Critical. ## Standards We Anchor On - OWASP Top 10:2025 (web) — https://owasp.org/Top10/2025/ - OWASP API Security Top 10:2023 — https://owasp.org/API-Security/editions/2023/en/0x11-t10/ - OWASP Top 10 for LLM Applications:2025 — https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/ - 2025 CWE Top 25 — https://cwe.mitre.org/top25/archive/2025/2025_cwe_top25.html - OWASP Web Security Testing Guide — https://owasp.org/www-project-web-security-testing-guide/ ## Open-Source Toolchain (transparency) bleek embeds the following permissively-licensed tools (CLI subprocess pattern for GPL/AGPL compatibility): - gitleaks (MIT) — secrets in bundles - trufflehog (AGPL) — secrets with live verification, subprocess-only - OWASP ZAP (Apache 2.0) — injection DAST - Nuclei (MIT, ProjectDiscovery) — CVE templates, exposure checks - sslyze (GPLv2) — TLS analysis, subprocess-only - Mozilla Observatory v2 API — header grading - Retire.js (Apache 2.0) — client-side JS dependency vulnerabilities - OSV.dev (Apache 2.0, Google) — vulnerability advisory data - Garak (Apache 2.0, NVIDIA) — LLM prompt-injection / system-prompt-leakage probes (Phase 2) Proprietary checks bleek built in-house: - Supabase RLS probing pipeline - Firebase rule probing pipeline - Vibe-coding-platform fingerprinting - AI-Ready Fixes Markdown schema - Trust badge issue + auto-revoke ## Trust Posture - Read-only by design. We never run destructive payloads. - Free and Deep scans need no code access. The URL is enough. - Human Checker tier ships fixes as pull requests the founder approves. - Named founder: Ali Saeed (Alenor brand). - NDA available on request. ## Founder Ali Saeed runs bleek under the Alenor brand. Background: production engineer fluent in React, Node, PostgreSQL, AI coding tools (Claude Code, Cursor), and the vibe-coding stack (Lovable, Bolt, Supabase, Vercel). No agency, no offshore team, no ghosting. ## Links - Website: https://bleek.dev - Free scan: https://bleek.dev/#scan - Pricing: https://bleek.dev/#pricing - Blog: https://bleek.dev/blog - Contact: hello@bleek.dev - Founder: Ali Saeed, Alenor