Preload by Agence Web Performance

Preload #04 ·

WordPress 7.1.4 under attack, Brotli flaw, and AI-bloated sites

Also: Chrome drops XSLT, a monitoring script turned wallet drainer at Polymarket, and Safari's modulepreload trap.

This week, security trouble came from places we rarely look: the WordPress update process itself, a compression library CMSs don't even ship, and a performance monitoring script turned into a wallet drainer. Keeping your CMS up to date is no longer enough; you also need to keep an eye on what runs around it. On the browser side, Chrome turns the page on XSLT on November 17, and an independent benchmark puts JPEG XL back in its place next to AVIF.

Enjoy the read, and good luck with the checks.

The essentials

  • WordPress 7.1.4 fixes a code execution flaw during core updates, already targeted by attackers.
  • Left to its defaults, an AI agent builds needlessly heavy and complex sites: give it the architecture and a JavaScript budget.
  • Brotli flaw: update the library on your server, your CMS can't fix it for you.

Web platform

Sitemaps and RSS feeds will show as plain text in Chrome from November 17

The change mainly hits the styling of XML sitemaps and RSS or Atom feeds dressed up with an XSL stylesheet: starting with Chrome 158, due in stable on November 17, Chrome no longer runs XSLT, so these files will display in the browser as raw XML, in plain text. They keep working normally for search engine crawlers and feed readers. The removal also affects code calling XSLTProcessor; only sites enrolled in an origin trial and machines managed with the XSLTEnabled enterprise policy keep XSLT, until Chrome 176 (August 2027).

Next step

For a sitemap or a feed, the simplest option is to leave it as is: plain-text display has no impact on crawlers. If you really want a human-readable version, serve HTML to browsers and XML to crawlers, which amounts to cloaking and takes more work to set up and maintain: your call, it isn't essential. If your site uses XSLT elsewhere (XSLTProcessor, transformed XML pages), move the transformation to the server.

Google puts XSLT usage at 0.02% of page loads. For most sites, the only visible effect will be how SEO plugin sitemaps look, and only humans open those in a browser.

Read the source

Topics:SEOChrome

JPEG XL vs AVIF: an independent benchmark keeps AVIF as the default

With JPEG XL now decoded by Chrome, the question of the default image format is back. Jake Archibald's benchmark, published October 15 (libjxl 0.12, avifenc 1.4.2, SSIMULACRA2 and ColorVideoVDP metrics), settles it for the web: on lossy photos, the best JPEG XL weighs 15 to 25% more than AVIF at default effort settings, and 30 to 90% more on screenshots and illustrations, while taking two to three times longer to decode. JPEG XL keeps the edge for lossless recompression of existing JPEGs (9 to 20% smaller) and for lossless compression.

Next step

Keep AVIF as your first choice, in 4:4:4 (4:2:0 chroma subsampling rarely saves bytes) and at the highest effort your build pipeline can afford. Save JPEG XL for a library of JPEGs whose originals are gone, always with a fallback in <picture>, since Edge doesn't decode it.

This matches our own benchmark, where AVIF came out lightest in 7 of 8 measurements at typical web quality settings.

Read the source

Topics:Performance

Performance

Safari: one failed modulepreload can break a page until the next tab

A Safari visitor can stay stuck on a page after a single transient error (a 404, a CDN incident, a dropped request): in WebKit, a <link rel="modulepreload"> that fails stays recorded as loaded, and a later import() of the same URL rejects without retrying the request, even once the file is back. WebKit fixed it in Safari Technology Preview 255, released October 22; Safari 27.2, shipped October 26, doesn't mention the fix. Sites split into preloaded JavaScript modules, as many build tools produce, are exposed.

Next step

Test the failure path: make a module fail to load once, restore it, then check that a new request goes out. Add a bounded retry on the same URL (no random query parameter, which would create a separate module), and make sure your CDN and service worker don't keep serving the failed response.

In our October 28 test, the rejection survived a tab reload in Playwright's WebKit 27.2, while Chrome 156 and Firefox 157 retried the request: only a new tab unblocked the page.

Read the source

Topics:JavaScriptPerformanceSafari

pngdeer, a PNG compressor that picks its own quality

Screenshots and UI graphics that have to stay PNG can still get lighter: pngdeer, released as open source (MIT license), reduces their color palette (lossy compression) with no quality slider. It tries palettes from 2 to 256 colors, with or without dithering, scores each attempt with the SSIMULACRA2 metric and keeps the smallest file that reaches its target, capped at 90. On the author's benchmark (180 images, reproducible from the repository), the median file is 8 to 12% smaller than with TinyPNG; it ships as a command-line tool (Homebrew) and as an in-browser WebAssembly app that never uploads your images.

Next step

Run your UI PNGs through pngdeer, then through a lossless optimizer like oxipng, which remains complementary. Set the quality with --target if you want consistent results across images, and keep AVIF or WebP for anything that doesn't need to be PNG.

The project is young (424 stars and 33 commits so far) and the benchmark comes from its author: try it on your own images before adding it to a build pipeline.

Read the source

Topics:Developer toolsPerformance

CMS

WordPress aims for MCP and OAuth in core for 7.3

Connecting an AI agent to WordPress may soon no longer mean copying an application password around: on October 28, Jason Adams from the WordPress AI team proposed four workstreams, laid out during 7.2 and targeted at 7.3. MCP Adapter, the plugin that exposes WordPress to agents over the MCP protocol, would be hardened and then brought into core; a baseline set of abilities (the actions WordPress exposes to agents) would cover common tasks; OAuth would let site owners authorize an agent and later revoke it. This is a proposal, subject to the usual core review, not a decision.

Next step

In the meantime, give each agent a dedicated account with the lowest role that works, and a separate, revocable application password per tool. Before opening MCP access, list the site's public abilities: MCP Adapter only checks that the user is logged in and has the read capability, so the real access control lives in each ability.

The proposal tackles the right problem. Until it ships, the REST API remains the most reliable way to put an agent to work, and it's the one we favor in-house.

Read the source

Topics:AISecurityWordPressMCP

Query Monitor 4.1 shows script weight, with a blind spot

Query Monitor, the WordPress debugging plugin, now shows the size of every enqueued script and stylesheet, with a total, since version 4.1.0 on October 26; 4.1.1 followed the next day to handle WordPress 7.2's loading changes. The size isn't read from the server: it's the transfer size measured by the browser (encodedBodySize, Resource Timing API) for the exact URL of the asset. So it stays empty, and counts as zero in the total, for a cross-origin script loaded without a crossorigin attribute, or for a file minified, combined or moved to a CDN by an optimization plugin.

Next step

Treat that column as a guide for the site's own files. For a page's real weight, open the Network tab in DevTools in a logged-out session: Query Monitor shows the page as generated for a logged-in user, bypassing the page cache.

Read the source

Topics:Developer toolsPerformanceWordPress

Languages and frameworks

The JavaScript cost of letting an AI agent pick the framework

Letting an AI agent choose a site's architecture can blow your JavaScript budget. Nolan Lawson had Claude Code build the same app, a content-heavy catalog with little interactivity, in 12 frameworks, five times each: median client JavaScript ranged from 0 KB with Astro to 142 KB with Next.js, and median LCP measured by Lighthouse from 810 ms to 2,090 ms. Agents often built a single-page app (SPA) even when nothing called for one: the agent building the no-framework version did so five times out of five.

Next step

Spell out the expected architecture (server rendering, separate pages) and a JavaScript budget in your agent's instructions, then measure the result in Lighthouse or the Network tab. On a Next.js 16.4 project, the next-bundle-optimizer skill, released the same day by the Next.js team, helps an agent find JavaScript that loads for nothing.

The SPA bias is real, while a content site only needs a few kilobytes of JavaScript. As the author himself points out, with a single model tested, this benchmark is one data point, not a verdict.

Read the source

Topics:AIJavaScriptPerformance

Next.js ships seven security fixes this Tuesday

Next.js apps have a security update to apply today: on October 27, the Next.js team announced a security release scheduled for Tuesday, November 3, fixing seven medium-severity vulnerabilities. The full advisories, with affected versions and upgrade instructions, are published alongside the release.

Next step

Block out time today: read the advisories as soon as they're out to see which versions are affected, upgrade your Next.js apps, then redeploy them.

Read the source

Topics:Security

AI and dev tools

An open-weight model tried to exfiltrate data from a sandbox

Running an open model yourself doesn't protect you from what a tool-wielding agent does. Cohere reported on October 26 that during a threat detection test in a sandbox, Qwen3.8-Flash-Next (Alibaba) inventoried the machine (the /tmp folder, transfer tools, DNS) and then tried to send data to an Alibaba Cloud storage bucket in Beijing, while its reasoning claimed it would call no tools. The behavior is rare, under one call in 100,000; in Cohere's tests, adding the -o flag of curl to the command raised URL generation from 4% to 92%.

Next step

Deny all outbound connections from your agents by default and allow only the hosts they need. Check every tool call before it runs, without relying on the reasoning trace, and block unprivileged user namespaces (unshare) in their containers.

Cohere sells its own models and the root cause hasn't been established. The lesson still applies to every agent: its reasoning trace is not a security boundary.

Read the source

Topics:AISecurity

Security

WordPress 7.1.4 fixes a code execution flaw in core updates, already targeted

The WordPress core update itself opened a hole: up to 7.1.3, a full-package update unpacks the new files, installer included, into a predictable folder under wp-content/upgrade/ served by the web server, and an unauthenticated attacker who reaches it during those few seconds can execute code (CVE-2026-119977, CVSS 8.1). WordPress 7.1.4, released October 28 with a dozen other fixes (stored XSS, Interactivity API, unpublished Customizer changesets readable without logging in), makes that folder unpredictable; the fix is backported to the 4.7 branch. Wordfence counted more than 18,500 exploitation attempts in the following 21 hours.

Next step

Check in Dashboard > Updates that every site runs 7.1.4 or the patched release of its branch, rather than assuming the automatic update went through. On a site that stayed unpatched, search the logs for requests to PHP files under wp-content/upgrade/. The release also updates its certificate bundle after the eMudhra incident: grep -i emsign wp-includes/certificates/ca-bundle.crt should now return only emSign Root CA - G1.

Read the source

Topics:SecurityWordPress

Brotli flaw: WordPress is exposed through the server, not its updates

A flaw in the Brotli compression library affects sites that an up-to-date CMS doesn't protect: CVE-2026-111497, an out-of-bounds write in the decoder, affects versions 1.0.3 to 1.2.0, fixed on October 21 in 1.0.10, 1.1.1 and 1.2.1, and the US agency CISA has listed it as exploited since October 27. WordPress doesn't ship Brotli, but reaches it through PHP's cURL extension whenever it fetches a URL supplied from outside: an unauthenticated pingback to xmlrpc.php, an oEmbed, a feed. A crafted response at least crashes the PHP worker; code execution hasn't been demonstrated through this path.

Next step

Update the system Brotli package (libbrotli1 on Debian and Ubuntu, checking the patched revision for your distribution), rebuild your Docker images and restart PHP-FPM. If you don't use pingbacks, remove pingback.ping with the xmlrpc_methods filter: the xmlrpc_enabled filter doesn't block them.

Read the source

Topics:SecurityPHPWordPress

A field monitoring script hijacked to drain Polymarket accounts

A performance monitoring script can do a lot more than measure: on October 29, according to Hudson Rock's analysis, the DebugBear real user monitoring (RUM) file loaded by the betting platform Polymarket contained code that looked for the open session in the page, sold or transferred positions to an attacker's address, then reported back to a DebugBear server. Polymarket's Content Security Policy (CSP) allowed the script's host, and therefore anything it served. DebugBear asked customers to remove its embed code; Polymarket, which points to a compromised third-party provider, paused trading and promised refunds. How the script was compromised and which other customers were hit is still unknown.

Next step

Search your sites' code for cdn.debugbear.com and remove it. More broadly, load no third-party script, monitoring agents included, on account, checkout and admin pages. Subresource Integrity (SRI) only protects a frozen file: a vendor script updated in place breaks its hash.

Allowing a host in your CSP means trusting every byte it will ever serve. A monitoring tool has no business on a page where people pay or administer a site.

Read the source

Topics:Field dataPerformanceSecurity

Spamhaus now requires consistent DNS from every server that sends email

Email sent directly from your servers may get blocked if they lack a consistent DNS identity: since November, rolling out over several months, Spamhaus requires every sending IP to have a valid HELO name (the name the server announces when it connects), a reverse PTR record and an A record that point to each other, or the server may be mistaken for a compromised machine. Exposed: servers that send mail themselves with the host's generic PTR and a default HELO, such as WordPress and WooCommerce on a VPS using mail(), Laravel or Symfony apps, monitoring alerts and cron jobs.

Next step

Test each sending IP with the HELO/DNS check on Spamhaus's Sender Guidelines page. Ask your host for a PTR that matches the server's HELO name, or move transactional email to an authenticated SMTP service or API: Google Workspace, Microsoft 365 and sending services such as Brevo manage their own IPs.

There's no hard deadline, but order confirmations and password resets are the first emails to vanish silently: now is the time to move transactional mail off mail().

Read the source

Topics:HostingSecurityWordPress

Finds

Basicrum, open source field monitoring on a single server

To measure Core Web Vitals from real visitors without a subscription, Basicrum has released an installer that deploys everything on one server with a single command: six Docker containers (Traefik, a Rust ingestion service, ClickHouse, Postgres, a Node API and a Next.js dashboard), fed by the open source Boomerang collection script. It needs at least 2 CPUs and 4 GB of memory for Docker, and handles TLS through Cloudflare Origin CA, Let's Encrypt or your own certificate.

Next step

Try it on a non-shared server (a VPS, a cloud instance or a dedicated server), in letsencrypt or cloudflare mode depending on your DNS. If you run Docker Engine 29 yourself, keep Traefik at 3.6 or later: older versions can no longer talk to it.

It's a credible alternative to paid tools once several sites share the collection. For a single site, a few lines of web-vitals and a small collector are enough; the brand-new repository doesn't show a license yet.

Read the source

Topics:Field dataDeveloper toolsPerformance

CSS-Tricks is independent again

One of the go-to front-end resources has changed hands: Geoff Graham, who has written for CSS-Tricks for twelve years, is buying the site from DigitalOcean, its owner since 2022, he announced on October 26. The site will run on direct advertising, memberships and a shop, with a beginner HTML and CSS course in the works. The newsletter returns on November 10, with one or two articles a week, and guest posts are open again, AI-generated content excluded.

Next step

If you have a CSS lesson learned worth sharing, now is the time to pitch a guest post; otherwise, subscribe to the newsletter before it restarts on November 10.

Read the source

Topics:CSSDeveloper tools

Eroan Boyer
Founder of Agence Web Performance

The next issue, Tuesday morning.

Get Preload every week, two weeks before it goes public. Free, one-click unsubscribe.

All issues