<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[ShipSniffer]]></title><description><![CDATA[ShipSniffer]]></description><link>https://shipsniffer.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Wed, 18 Nov 2026 14:58:08 GMT</lastBuildDate><atom:link href="https://shipsniffer.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[The boring launch checklist: what 99 vibe-coded apps forgot before going live]]></title><description><![CDATA[Real scan data from 99 live apps shared on Hacker News]]></description><link>https://shipsniffer.hashnode.dev/boring-launch-checklist-vibe-coded-apps</link><guid isPermaLink="true">https://shipsniffer.hashnode.dev/boring-launch-checklist-vibe-coded-apps</guid><category><![CDATA[Security]]></category><category><![CDATA[webdev]]></category><category><![CDATA[vibe coding]]></category><category><![CDATA[SEO]]></category><category><![CDATA[launch]]></category><dc:creator><![CDATA[ShipSniffer]]></dc:creator><pubDate>Tue, 17 Nov 2026 15:07:20 GMT</pubDate><content:encoded><![CDATA[<p><em>November 2026 · Real scan data from 99 live apps shared on Hacker News</em></p>
<p>Last week we wrote about leaked API keys, the scary failure. This week is about the boring ones. They won't empty your OpenAI balance overnight. They quietly make your app look unfinished to browsers, search engines, and anyone who pastes your link into a chat.</p>
<p>In early November 2026 we pointed the ShipSniffer scanner at apps people had linked in Hacker News posts. Many were built with Lovable, Replit, Bolt, or Vercel templates. We tried 155 URLs. 56 were dead, blocked, or returned an error, which leaves <strong>99 live apps</strong> with a full scan. Every number below comes from those 99 scans. We made only ordinary requests for public pages, and we aren't naming any of the apps.</p>
<p>Where they were hosted:</p>
<table>
<thead>
<tr>
<th>Host</th>
<th>Apps</th>
</tr>
</thead>
<tbody><tr>
<td>lovable.app</td>
<td>45</td>
</tr>
<tr>
<td>Custom domain</td>
<td>34</td>
</tr>
<tr>
<td>replit.app</td>
<td>11</td>
</tr>
<tr>
<td>vercel.app</td>
<td>6</td>
</tr>
<tr>
<td>netlify.app</td>
<td>2</td>
</tr>
<tr>
<td>Bolt</td>
<td>1</td>
</tr>
</tbody></table>
<p>The good news first: these apps mostly weren't disasters. The median ShipSniffer score was <strong>84/100</strong> and the average was <strong>82.7</strong>. Grades came out as 13 A's, 68 B's, 16 C's, and 2 D's. Most of the points were lost on things a dev can fix in an afternoon.</p>
<h2>1. Security headers: almost everyone skipped them</h2>
<p>Security headers are lines your server sends with every page. They tell the browser what the page may do. AI app builders rarely add them, because nothing breaks without them.</p>
<p>Out of 99 apps:</p>
<ul>
<li><strong>90 (91%)</strong> had no Content-Security-Policy</li>
<li><strong>87 (88%)</strong> had no X-Frame-Options or <code>frame-ancestors</code>, so anyone could load them inside an iframe on another site</li>
<li><strong>38 (38%)</strong> had no Referrer-Policy</li>
<li><strong>37 (37%)</strong> had no X-Content-Type-Options</li>
<li><strong>8 (8%)</strong> had no Strict-Transport-Security (HSTS)</li>
<li><strong>5 (5%)</strong> didn't pass a single header check</li>
</ul>
<p>HSTS was the one most apps got right, mostly because hosts like Vercel and Netlify send it for you. The rest are up to you.</p>
<p><strong>The fix.</strong> On Vercel, add a <code>headers</code> block to <code>vercel.json</code>. On Netlify, use a <code>_headers</code> file. On most other hosts you set them in the server or framework config. A safe starting set:</p>
<pre><code>X-Content-Type-Options: nosniff
Referrer-Policy: strict-origin-when-cross-origin
X-Frame-Options: DENY
Content-Security-Policy: frame-ancestors 'none'
</code></pre>
<p>A full Content-Security-Policy that lists every script source takes more care. A bad one can break your Supabase or Stripe calls. Start with <code>frame-ancestors</code>, then tighten it over time. You can also ship the full policy as <code>Content-Security-Policy-Report-Only</code> first, so you can watch what it would block before turning it on.</p>
<h2>2. Link previews: 30% looked blank when shared</h2>
<p>When you drop a link in Slack, Discord, X, or iMessage, the preview card is built from Open Graph tags. <strong>30 of 99 apps (30%)</strong> were missing Open Graph tags. <strong>11 (11%)</strong> had no meta description at all.</p>
<p>Launch day is when your link gets shared the most. A bare URL with no title, image, or blurb gets fewer clicks than one with a clean card.</p>
<p><strong>The fix.</strong> Put these in your <code>&lt;head&gt;</code>. In a Vite or React SPA that's <code>index.html</code>. In Next.js, use the <code>metadata</code> export:</p>
<pre><code class="language-html">&lt;title&gt;YourApp – one-line promise&lt;/title&gt;
&lt;meta name="description" content="What it does, for whom, in one sentence."&gt;
&lt;meta property="og:title" content="YourApp – one-line promise"&gt;
&lt;meta property="og:description" content="Same sentence, or a punchier one."&gt;
&lt;meta property="og:image" content="https://yourapp.com/og.png"&gt;
&lt;meta property="og:url" content="https://yourapp.com/"&gt;
&lt;meta name="twitter:card" content="summary_large_image"&gt;
</code></pre>
<p>Use an absolute URL for <code>og:image</code>. A 1200×630 PNG is the safe size. Then paste your link into a private chat to check the card before you post it anywhere public.</p>
<h2>3. Sitemaps: 77% didn't have one</h2>
<p><strong>76 of 99 apps (77%)</strong> returned nothing at <code>/sitemap.xml</code>. For a single-page app with one URL, that's not a disaster. But once you add a pricing page, docs, a blog, or public profile pages, a sitemap is the easiest way to help search engines find them.</p>
<p><strong>The fix.</strong> For a small site, a hand-written <code>public/sitemap.xml</code> with your main URLs is enough. Next.js can generate one from <code>app/sitemap.ts</code>. Then add a <code>Sitemap: https://yourapp.com/sitemap.xml</code> line to <code>robots.txt</code>.</p>
<p>While you're in <code>robots.txt</code>, check that you aren't blocking crawlers by accident. Some templates ship a blanket <code>Disallow: /</code> that was only meant for staging.</p>
<h2>4. Telling the world what you run on</h2>
<p><strong>13 apps (13%)</strong> sent an <code>X-Powered-By</code> header announcing their server framework. It's a small leak. It won't get you hacked by itself, but it gives a free hint to anyone hunting for apps on a version with a known bug. In Express, <code>app.disable('x-powered-by')</code> removes it. Next.js has a <code>poweredByHeader: false</code> option.</p>
<p><strong>2 apps (2%)</strong> served public source maps. Source maps rebuild your original source code in the browser's dev tools. That's handy for debugging, but in production it hands out your readable code, including any comments or logic you'd rather keep private. Turn them off for production builds (for example, <code>build.sourcemap: false</code> in Vite, which is the default) unless you upload them privately to an error tracker.</p>
<p><strong>3 apps (3%)</strong> had broken internal links. These are usually leftover nav items pointing to pages the AI planned but never built.</p>
<h2>5. The scary stuff was rare, but not zero</h2>
<p>In this set of 99, <strong>1 app</strong> exposed a live OpenAI secret key in its client code. We reported it privately and won't share anything that identifies it. <strong>4 apps (4%)</strong> had a Google API key in client code. Some Google keys are designed to be public, like Maps keys, but only when they're restricted to your domain and the APIs you actually use in the Google Cloud console.</p>
<p>No app in this set exposed a <code>.env</code> file or a <code>.git</code> folder at its public URL. That's good to see, and it's still worth checking yours.</p>
<p>If you build on a static host and call an AI API, our <a href="https://shipsniffer.hashnode.dev">leaked-keys write-up</a> covers why the key has to move to a server or edge function.</p>
<h2>A 15-minute pre-launch pass</h2>
<p>Before you post your app anywhere, run through this:</p>
<ol>
<li>Add the four starter security headers above.</li>
<li>Add a title, meta description, and Open Graph tags, then test the preview card in a private chat.</li>
<li>Add <code>sitemap.xml</code> and check that <code>robots.txt</code> isn't blocking everything.</li>
<li>Remove <code>X-Powered-By</code> and turn off public source maps.</li>
<li>Click every link in your nav.</li>
<li>Search your built JavaScript for <code>sk-</code>, <code>sk-proj-</code>, and <code>sk-ant-</code>. If you find a secret key, revoke it and move the call to the server.</li>
</ol>
<p>None of this is glamorous. All of it decides whether your launch looks finished or looks like a weekend demo.</p>
<hr />
<p><em>Want the checklist run for you? <a href="https://shipsniffer.com">ShipSniffer</a> gives your live URL a free scan for security headers, meta tags, sitemaps, leaked keys, and more in about a minute, with plain-English fixes. It only uses ordinary public requests.</em></p>
<p><em>Method: ShipSniffer's scanner checked 155 app URLs linked from Hacker News posts in early November 2026. 99 returned a live page and are counted above. Percentages are out of those 99. Sites are not named.</em></p>
]]></content:encoded></item><item><title><![CDATA[What we found scanning vibe-coded apps for leaked API keys]]></title><description><![CDATA[November 2026 · A look at 157 confirmed public leaks Shipping an app can take an afternoon: a prompt, a static host, and a live URL. The cost of that speed…]]></description><link>https://shipsniffer.hashnode.dev/what-we-found-scanning-vibe-coded-apps-for-leaked-api-keys</link><guid isPermaLink="true">https://shipsniffer.hashnode.dev/what-we-found-scanning-vibe-coded-apps-for-leaked-api-keys</guid><category><![CDATA[Security]]></category><category><![CDATA[vibe coding]]></category><category><![CDATA[webdev]]></category><category><![CDATA[openai]]></category><dc:creator><![CDATA[ShipSniffer]]></dc:creator><pubDate>Fri, 13 Nov 2026 14:19:30 GMT</pubDate><content:encoded><![CDATA[<p><em>November 2026 · A look at 157 confirmed public leaks</em></p>
<p>Shipping an app can take an afternoon: a prompt, a static host, and a live URL. The cost of that speed shows up in a predictable place: secret API keys baked into pages anyone can open.</p>
<p>Between October 21 and November 13, 2026, ShipSniffer ran a weekday hunt for live secret keys on public, vibe-coded sites. We only made ordinary GET requests for pages and same-site JavaScript. We never used a key, never called a provider with a finding, and never published a hostname, repo, email, or key value. When a site still had a live secret and we could find a contact, we sent a private disclosure.</p>
<p>In that window we confirmed live secret API keys on <strong>157</strong> publicly reachable sites.</p>
<h2>Almost all of them were OpenAI keys</h2>
<p>Of those 157 sites:</p>
<ul>
<li><strong>154 (98%)</strong> exposed an OpenAI secret key</li>
<li><strong>3 (2%)</strong> exposed an Anthropic secret key</li>
</ul>
<p>We did not confirm a live Stripe, AWS, or GitHub token in this set. That is not a claim those leaks do not exist; it is what this hunt found.</p>
<p><strong>145 of 157 sites (92%)</strong> clearly used OpenAI's project-key format (<code>sk-proj-</code>). Three were Anthropic keys (<code>sk-ant-</code>). The remaining nine were OpenAI keys in an older or unspecified <code>sk-</code> format. A project key in your dashboard is still a secret. If it ships to the browser, anyone who views source or a JavaScript bundle can spend against your account.</p>
<h2>GitHub Pages was the main stage</h2>
<p>The 157 sites broke down by host platform like this:</p>
<table>
<thead>
<tr>
<th>Platform</th>
<th>Sites</th>
<th>Share</th>
</tr>
</thead>
<tbody><tr>
<td>GitHub Pages</td>
<td>135</td>
<td>86%</td>
</tr>
<tr>
<td>Vercel</td>
<td>17</td>
<td>11%</td>
</tr>
<tr>
<td>Netlify</td>
<td>2</td>
<td>1%</td>
</tr>
<tr>
<td>Lovable</td>
<td>1</td>
<td>&lt;1%</td>
</tr>
<tr>
<td>Custom domain or other</td>
<td>2</td>
<td>1%</td>
</tr>
</tbody></table>
<p>GitHub Pages is a great free host for a demo. It is also static. There is no server-side environment variable and no private backend unless you add one. If the app calls OpenAI, the only place a key can live is in the files Pages serves to the world. That is exactly where we found them.</p>
<p>Vercel showed up less often, which matches its defaults: secrets belong in server or edge routes, not the client bundle. The 17 Vercel sites in this set still put a secret in code the browser downloads.</p>
<h2>The keys were not hidden. They were in <code>index.html</code> and <code>script.js</code>.</h2>
<p>We classified each finding by the file it appeared in:</p>
<ul>
<li><strong>84 of 157 (54%)</strong> were in an HTML file, including <strong>45 (29%)</strong> sitting in <code>index.html</code></li>
<li><strong>58 of 157 (37%)</strong> were in a JavaScript file, including <strong>40 (26%)</strong> in a file named <code>script.js</code></li>
<li><strong>15 of 157 (10%)</strong> were in the page or an inline script without a clean filename in our notes; most of those were homepage HTML</li>
</ul>
<p>On GitHub Pages the pattern was almost mechanical: a static <code>index.html</code>, sometimes a <code>script.js</code> next to it, and a key assigned as a constant near the app config. A smaller set lived in hashed production bundles (Vite or webpack), which is no safer. A hashed filename is not access control. If you can "View Source" and see a key, everyone else can too.</p>
<h2>Why this keeps happening</h2>
<p>The usual path is not a sophisticated attacker or a compromised host:</p>
<ol>
<li>An AI coding tool or tutorial shows a working <code>fetch()</code> to the OpenAI or Anthropic API.</li>
<li>The key is pasted into the frontend so the demo runs.</li>
<li>The project is published to GitHub Pages or a similar static host.</li>
<li>The deploy succeeds, the chat UI works, and the key stays in the shipped files.</li>
</ol>
<p>Nothing in that flow looks broken from the builder's chair. The app responds. The problem is that the "backend" is running in every visitor's browser.</p>
<h2>What to do if this is your app</h2>
<p>Do these in order. Deleting the key from the repo and redeploying is not enough, because the old key is already burned.</p>
<p><strong>1. Rotate the key immediately.</strong> In the OpenAI or Anthropic dashboard, revoke the exposed key and create a new one. Treat the old value as public, even if you think the site had little traffic.</p>
<p><strong>2. Set a spend limit.</strong> Turn on usage caps and billing alerts on the provider account before you do anything else with a new key. A leaked key with no cap can run a bill overnight. A hard monthly limit (even a small one) contains the damage.</p>
<p><strong>3. Move the call off the client.</strong> The browser should talk only to your own backend. Put the provider call in a server route, serverless function, or edge function. The new key never ships in HTML or a client bundle.</p>
<p><strong>4. Store the new key in an environment variable</strong> that only the server can read. Do not prefix it with <code>NEXT_PUBLIC_</code>, <code>VITE_</code>, or <code>REACT_APP_</code>. Those prefixes exist to embed values in the client build.</p>
<p><strong>5. Remember that GitHub Pages cannot keep this secret.</strong> A static Pages demo needs a tiny backend somewhere else, or it should stop calling the provider. There is no Pages setting that hides a key in <code>index.html</code>.</p>
<p><strong>6. Check git history, not just the latest commit.</strong> Removing the key from <code>main</code> does not remove it from older commits, GitHub's file history, or cached deploys. After you rotate, search history and purge any still-public copies.</p>
<p>A short test for the obvious cases: open the live site, view source, and search the page and downloaded JavaScript for <code>sk-proj-</code>, <code>sk-ant-</code>, and <code>sk-</code>. If it is there, visitors already have it.</p>
<h2>We notified owners when we could</h2>
<p>Of the 157 confirmed sites, we sent <strong>85</strong> private disclosure notices. For the other <strong>72</strong>, we could not find a contact email, so nothing was sent. Almost every notice used the email on a git commit in the public repo. If you ship a public demo, a reachable contact is how someone can tell you a key is live.</p>
<p>We are not naming the sites, and we will not post key fragments. The useful part of this data is the pattern, not a hall of shame.</p>
<h2>Check your own app</h2>
<p>If you have shipped a vibe-coded demo, especially on GitHub Pages, assume a secret in the client is public and go through the steps above. It takes a few minutes and is cheaper than a surprise invoice.</p>
<blockquote>
<p><strong>Run a free scan.</strong> Paste your URL into ShipSniffer for a passive check: leaked-key detection, missing headers, and copy-paste fix prompts. We do not test keys and we do not store the full value of a secret.</p>
<p><a href="https://shipsniffer.com/?ref=hashnode">Scan your app free →</a></p>
</blockquote>
<h2>Method</h2>
<p>Window: October 21 through November 13, 2026. We counted a site only if a secret-shaped key was still live in a public page or same-site JavaScript file when we checked it. Five false positives and two sites that were no longer live were excluded, which is why the confirmed count is 157.</p>
<p>Each check was a single passive pass: GET requests only. We never authenticated, never called a provider with a found key, and never tried to access an account. Figures are exact for our confirmed log, not a census of the whole web. Automated scans can miss keys and can flag lookalikes.</p>
]]></content:encoded></item><item><title><![CDATA[We scanned 155 launched vibe-coded apps. Here's what we found.]]></title><description><![CDATA[State of vibe-coded app security · October 2026 · 5 min read Vibe coding made shipping easy. Lovable, Bolt, Replit and Cursor take you from prompt to live…]]></description><link>https://shipsniffer.hashnode.dev/we-scanned-155-launched-vibe-coded-apps-here-s-what-we-found</link><guid isPermaLink="true">https://shipsniffer.hashnode.dev/we-scanned-155-launched-vibe-coded-apps-here-s-what-we-found</guid><category><![CDATA[Security]]></category><category><![CDATA[webdev]]></category><category><![CDATA[AI]]></category><category><![CDATA[supabase]]></category><category><![CDATA[vibe coding]]></category><dc:creator><![CDATA[ShipSniffer]]></dc:creator><pubDate>Sat, 17 Oct 2026 16:11:03 GMT</pubDate><content:encoded><![CDATA[<p><em>State of vibe-coded app security · October 2026 · 5 min read</em></p>
<p>Vibe coding made shipping easy. Lovable, Bolt, Replit and Cursor take you from prompt to live URL in an afternoon. We wanted to know what that URL looks like from the outside, so we collected 155 apps that their makers had publicly launched on Show HN (mostly <code>*.lovable.app</code>, <code>*.replit.app</code>, Vercel, Netlify and Bolt, plus apps on custom domains) and ran ShipSniffer's free passive scan on each one.</p>
<p>99 were still live and scannable. Here is what they had in common.</p>
<ul>
<li><strong>91%</strong> ship with no Content-Security-Policy</li>
<li><strong>88%</strong> can be framed by any site (no clickjacking protection)</li>
<li><strong>37%</strong> expose a Supabase anon key, so all their data security rides on RLS</li>
<li><strong>36%</strong> of launched platform-hosted links are already dead (404)</li>
</ul>
<h2>The good news: almost nobody leaked a secret key</h2>
<p>We looked through each app's HTML and up to 25 same-site JavaScript bundles for secret-shaped strings: OpenAI, Anthropic, Stripe live, AWS, GitHub, Slack and SendGrid keys, private key blocks, and Supabase <code>service_role</code> JWTs. <strong>Only 1 of 99 apps (1%) shipped a secret key to the browser</strong>, in its case an OpenAI key. That's much better than the horror stories suggest. The platforms' "put secrets in edge functions" defaults seem to be working.</p>
<p>But one is still enough: anyone who opens DevTools on that app can run up its OpenAI bill. We aren't naming it, and we never publish key values.</p>
<p>No app exposed <code>/.env</code> or <code>/.git</code>, and only 2% served public JavaScript source maps.</p>
<h2>The real gap: Supabase without proof of RLS</h2>
<p>37% of the live apps ship a Supabase anon key in their frontend. That's normal and by design. But it means one thing protects every row of user data: <strong>Row Level Security policies</strong>. One table with RLS switched off, or a policy of <code>using (true)</code>, and anyone with that public key can read the whole table.</p>
<p>A passive scan can't tell whether RLS is on, and we deliberately did <em>not</em> test anyone's database. (ShipSniffer only runs those checks on sites whose owners verify they own them.) If you're one of the 37%, open the Supabase dashboard and check every table in <code>public</code> before you launch. Firebase showed up in 4% of apps, and the same rule applies to its security rules.</p>
<h2>Security headers: the step everyone skips</h2>
<table>
<thead>
<tr>
<th>Header</th>
<th>Missing</th>
</tr>
</thead>
<tbody><tr>
<td>Content-Security-Policy</td>
<td>91%</td>
</tr>
<tr>
<td>X-Frame-Options / frame-ancestors</td>
<td>88%</td>
</tr>
<tr>
<td>Referrer-Policy</td>
<td>38%</td>
</tr>
<tr>
<td>X-Content-Type-Options</td>
<td>37%</td>
</tr>
<tr>
<td>Strict-Transport-Security</td>
<td>8%</td>
</tr>
<tr>
<td>All of the above</td>
<td>5%</td>
</tr>
</tbody></table>
<p>Hosting platforms add HSTS for you, so it's mostly fine. A CSP is the header that limits the damage when an AI-written component renders untrusted HTML, and almost no one has one. 13% also announce their stack with an <code>X-Powered-By</code> header.</p>
<h2>Launch polish</h2>
<p>30% were missing Open Graph tags (so shared links show up bare on X and Discord), 11% had no meta description, and 77% had no <code>sitemap.xml</code>. Overall grades: <strong>13% A, 69% B, 16% C, 2% D</strong>, with a median score of 84/100.</p>
<h2>The 5-minute pre-launch checklist</h2>
<ol>
<li>Search your built bundle for <code>sk-</code>, <code>sk_live_</code>, <code>service_role</code>. Rotate anything you find and move it into a server or edge function.</li>
<li>Supabase: turn on RLS for every table and remove any <code>using (true)</code> policy. Firebase: no test-mode rules.</li>
<li>Add a CSP and <code>frame-ancestors 'none'</code>. Start with <code>Content-Security-Policy-Report-Only</code> if you're nervous.</li>
<li>Turn off production source maps, or don't upload them publicly.</li>
<li>Add OG tags, a meta description and a sitemap so your launch post actually gets clicks.</li>
</ol>
<blockquote>
<p><strong>Check your app in about 20 seconds, free.</strong> Paste a URL to get a score, leaked-key detection, and copy-paste fix prompts for Cursor or Lovable. Apps that score A or B get a free "Scanned by ShipSniffer" badge.</p>
<p>👉 <a href="https://shipsniffer.com/?ref=hashnode">Scan my app free →</a></p>
</blockquote>
<h2>Method</h2>
<p>Sample: 155 URLs from Show HN posts that mention Lovable, Bolt, Replit, v0, Cursor or "vibe coded", collected in October 2026. 99 were reachable. 56 were dead, blocked or erroring, including 41 platform-hosted apps returning 404. Each app got a single passive pass: GET requests for the homepage, up to 25 same-site JS bundles, robots/sitemap/favicon, up to 30 internal links, plus <code>/.env</code>, <code>/.git/HEAD</code> and <code>.map</code> files for the first three bundles, at low concurrency. No logins, no database queries, no exploitation. Results are anonymized and reported only in aggregate. Automated scans produce false positives and negatives; treat these numbers as indicative.</p>
<p><em>Originally published at <a href="https://shipsniffer.com/blog/vibe-coded-app-security-2026">shipsniffer.com</a>.</em></p>
]]></content:encoded></item></channel></rss>