Show HN: I've made a CVE auto pentest app Keybrum launched a Show HN automated penetration-testing app that attacks live apps built with Lovable, Bolt, Replit, v0, Cursor and Claude Code, proving data leaks and exposed keys and handing fixes to coding agents for retest. Keybrum reports 576 individual checks across 339 attacks, a full run in about 1 minute, and cites CVE-2025-48757, noting 170 of 1,645 Lovable apps tested in 2025 let anyone read their database. A replayed run on Keybrum's demo app Pawpal executed 146 attacks in 206 requests, 20 of which worked, and the tool connects via an MCP server so an agent can read a finding, apply the fix and call the retest itself. CVE-2025-487571 in 10 Lovable apps leaked their data1 in 10 Lovable apps tested let anyone read their database https://nvd.nist.gov/vuln/detail/CVE-2025-48757 Hack your vibe-coded app before attackers do. An automated penetration test for apps built with Lovable, Bolt, Replit, v0 and Cursor. Keybrum attacks your live app like a hacker would, proves every data leak and exposed key, hands the fix to your coding agent, and attacks again until the hole is closed. or continue with Google https://keybrum.com/api/auth/google?next=%2Fapps or GitHub https://keybrum.com/api/auth/github?next=%2Fapps Free preview, no accountOnly apps you prove you own Attack run on pawpal.example 0 of 146 attacks · 0 worked 1. Visit 2. Read the code 3. Map the backend 4. Attack 5. Test keys 6. Report - Findings appear here once an attack works. A real run on Pawpal, our demo app, replayed: 146 attacks in 206 requests, 20 worked. Attacks apps built with these tools, on these backends - Lovable - Bolt - Replit - v0 - Cursor - Claude Code - Supabase - Firebase - Next.js - Express - tRPC - GraphQL AI builders ship a working app in an afternoon. Nobody checks what an attacker can do with it. So we built the attacker. - 576 - individual checks in 339 attacks, each with why it is safe on a live app - ~1 min - for a full run, while you watch it live - 0 rows - written by the write tests: they prove access without changing data - 170 / 1,645 - Lovable apps tested in 2025 let anyone read their database. CVE-2025-48757 https://nvd.nist.gov/vuln/detail/CVE-2025-48757 Keybrum attacks. Your agent fixes. The retest proves it. Keybrum finds what an attacker can reach, hands the proof and the fix to the coding agent you already use, and attacks again until the exploit fails. From found to fixed, on the live app. Connect the MCP server and your agent reads the finding, applies the fix and calls the retest itself. A finding closes only when the same attack fails. Retests are unlimited on every plan, and open for a week after a report. Below: the real session that fixed Pawpal's bookings finding, replayed. One loop, from the first attack to proof you can share. Start with one report on one app, or a plan for all of them. Every part works from the same proof: the request that worked, and the retest that failed. Proof of what an attacker can do. Keybrum loads your app like a visitor, reads its JavaScript for keys, maps the backend, then attacks with no account and as a brand-new one. No proof, no finding: you never sort through maybes. Anyone can read, add and change rows in profiles : 1,240 rows, with email, phone and name the public key shipped in your JavaScript, no account GET /rest/v1/profiles?select= &limit=10 apikey: eyJhbGciOiJIUzI1NiIsInR5cCI6… A fix written for the agent you already use. Each finding ships a fix prompt for Lovable, Bolt, Cursor or Claude Code, or the SQL. A leaked key also gets rotation steps: deleting it from the code does not un-leak it. Try the button: it copies the real prompt. CriticalAnyone can read, add and change rows in profiles : 1,240 rows, with email, phone and name Every deploy gets attacked again. A new publish is noticed and attacked again, with nothing to connect, on Lovable and Bolt too. Connect Vercel, Netlify or GitHub and previews are attacked before they ship: a CI check can stop the merge, and your agent can run the same check itself. You hear about it only when a new exposure is proven. When a new attack joins the library, every app is retested. - keybrum / attack the previewCame back: any signed-up user can read other people's bookings - buildSuccessful in 1m 12s - typecheckSuccessful in 31s Every new CVE, matched against what you run. Advisories are published every day, and nobody reads them all. Keybrum keeps your lockfile and the versions your live app serves, and matches each new advisory within the hour. Most pass you by. When one reaches you, your coding agent gets the upgrade already decided: the package, the version, the command. It runs your tests and ships, then Keybrum reads your live app and closes the advisory once the fixed version is there. 44 published advisories reach what it runs. 3 upgrades clear them all. - Criticalnext15.2.2 → 15.5.2837 advisoriesCVE-2025-55182Used in attacksProven on the live app - Highsharp0.33.5 → 0.35.53 advisories - Highpostcss8.4.31 → 8.5.234 advisories Published for its packages, and passed it by - 17 NovNext.js: Cache leak across root param values in nested 'use cache' functions via delayed inner stream metadataReaches versions from 16.2.0 until the fix in 16.4.5. It runs 15.2.2. - 16 NovNext.js: Information disclosure across root params when a 'use cache' function is called through unstable cache Reaches versions from 16.3.0 until the fix in 16.4.5. It runs 15.2.2. - 22 OctNext.js: Unauthenticated Remote Code Execution in use cache functions returning typed arraysReaches versions from 16.0.0 until the fix in 16.4.1. It runs 15.2.2. All the apps your team ships, ranked by what they expose. For agencies and companies whose staff build apps. Connect Vercel, Netlify or Supabase and every deployed app shows up, including the ones nobody declared. Each gets an owner, and reports carry your brand. The whole attack library, in the open. 339 attacks and 576 individual checks, each with what it tries and why it is safe to run against a live app. When one is added or improved, every app is retested against it. Plus, in your code: 62 classes of weakness. A scan file finds committed keys, open database rules and injectable queries in seconds, in your repository or your pipeline. For the rest, your own coding agent audits the code with Cloudflare's open-source security-audit skill https://github.com/cloudflare/security-audit-skill , handed over by Keybrum's MCP server. Your code stays on your machine; Keybrum receives the findings and proves what it can on the live app. How it works https://keybrum.com/docs code-audit - Core classes, in every codebase · 19 - State machine violations · Race conditions with business impact · Numeric/quantity manipulation · Access boundary violations · Implicit trust assumptions · Time-based logic · Default and fallback behavior · Export/backup as exfiltration · Import/restore as injection · Search/filter/sort as oracle · Enumeration through side effects · Preview/draft/staging leakage · Notification/webhook as SSRF · Multi-step boundary failures · Cross-component trust gaps · Second-order use · Scope and capability growth · Timing and ordering · Rollback and recovery - AI, LLM, and Agent · 3 - Context, retrieval, and memory · Tool and action · Output and disclosure - Client-Side and Browser · 5 - DOM and object-state · Cross-origin messaging and network · Service-worker and browser-storage · Cross-site information leaks · UI-redress and navigation - Cloud and Deployment · 5 - Workload identity and IAM · Ingress, network, and control-plane · Container and orchestration · Configuration and secret lifecycle · Managed storage, events, and edge - Data Isolation and Lifecycle · 4 - Tenant and object-isolation · Derived-data and disclosure · Export, backup, restore, and migration · Deletion, revocation, and lifecycle - Desktop, Mobile, and Local IPC · 5 - Deep-link, callback, and navigation · Webview and native-bridge · Local IPC and exported-component · Privileged-helper and local-file · Application-state and device-lifecycle - Memory Safety, Binary, and Kernel · 5 - Bounds, integer, and representation · Lifetime, type, and concurrency · FFI and ABI · Binary loading and runtime · Kernel and privileged-interface - Protocols, RPC, and Messaging · 4 - Framing, schema, and interpretation · RPC identity and authorization · Broker and queue isolation · Replay, ordering, and transaction - Resource Exhaustion and Availability · 4 - Computational amplification · Resource accumulation · Quota and scheduling · Failure and recovery - Supply Chain and Release · 3 - Dependency and build-input · CI and automation · Release and update - HTTP-Protocol and Authentication · 5 - HTTP framing and cache · Browser-session · Federated-identity · MFA, passkey, and account-transition · API-key and mTLS Attacks like a stranger. Behaves like a guest. Keybrum attacks only the apps you prove you own, and leaves with proof, not with your users’ data. Everything Keybrum does, and never does https://keybrum.com/security - Only apps you own - You prove the app is yours first: a tag, a DNS record added for you on Cloudflare , or a published file. Ownership is checked again every 30 days. - Proof without the data - Keybrum keeps counts, column names and one redacted row. It never stores the data it reads. - Write tests that write nothing - Strangers' writes are tested with a null primary key: the security rules run first, so the error proves they let us in, and no row is written. - Leaked keys, checked gently - A leaked key is tested with a free, read-only call, such as listing the models it can use. Then you get the steps to rotate it. - Light on your app - The Pawpal run: 146 attacks in 206 requests. Never a load test, never brute force. - Never your private network - Hosted Keybrum never attacks private addresses, and delete access is never tested. Built for whoever shipped the app. Founder, agency or team. For founders Find out what a stranger can do before your users do, and fix it with the agent you built the app with. Check my app https://keybrum.com/check For agencies Hand every client a third-party test at delivery, a trust page under your brand, and a retest on every deploy. See agency plans https://keybrum.com/pricing For companies Find every app your staff shipped, including the undeclared ones, ranked by what they expose, each with an owner. Talk to us https://keybrum.com/contact?topic=business Works where your app already lives Keybrum needs only the URL. Connect the rest when you want deploy checks, alerts or your agent in the loop. Pay once for a report. Or keep it attacked. A report attacks one app once: every finding with proof, the fixes for your agent, the PDF. A plan attacks every deploy again, keeps retests open and adds the trust page. The preview is always free. For a launch, or before you hand an app to a client. - 1 app, 1 full attack run - Proof for every finding - Fix prompts and MCP for your coding agent - Retests for 7 days to confirm fixes - PDF report of the findings For a founder shipping a few apps. - 5 apps, 2 members - Every deploy and publish attacked, CI check - Every new CVE matched against your dependencies - Attacked again every week or month - Unlimited retests - Trust page, badge, Slack and Discord alerts For agencies and small teams. - 25 apps, 10 members - Everything in Maker - Your brand on reports and trust pages - Every app deployed on Vercel, Netlify or Supabase found See a full sample report https://keybrum.com/keybrum-sample-report.pdf PDF, on our demo app . More than 25 apps, or every app your company builds? Talk to us https://keybrum.com/contact?topic=business Paid through Polar, with a VAT invoice. Plans change or cancel any time from Settings. Compare plans and billing questions https://keybrum.com/pricing Questions, before you paste the URL. Is this a code scanner? How is it different from my builder's security scan? Which apps can it attack? Does it watch my dependencies for new CVEs? Check a lockfile https://keybrum.com/tools/dependency-vulnerability-checker What if no attack works? Do you need my code or my database password? Can an attack break my app? How the attacks stay safe safety