WordPress has become a liability for your business.

WordPress is a content management system (CMS): software that lets you build and edit a website without writing code. It runs about 39.7% of all websites on the internet and roughly 58% of sites that use a known CMS.¹⁰

What made it famous was plugins. Need a contact form? Install a plugin. A photo gallery, fancy buttons, an online store, SEO tools, a booking calendar? There's a plugin for each. Someone with no coding background could assemble a professional-looking site in an afternoon, and a typical business site ends up running a dozen or more plugins, each from a different developer.

That convenience is exactly the problem.

WordPress plugins are not fenced off from each other or from the rest of your site. A plugin runs in the same space as WordPress itself, with direct access to your database and your files. A simple contact form plugin can, in principle, read unpublished pages, change another plugin, or send your data anywhere.¹¹¹

"WordPress' plugin architecture is fundamentally insecure."

Cloudflare, introducing its WordPress alternative, April 2026¹

The founder of Yoast SEO, one of the most widely used WordPress plugins ever made, makes the same point about plugins running with the full permissions of the site: "It's a security debt that compounds with every plugin install."¹²

The numbers back it up. Of the 11,334 new security flaws found in the WordPress ecosystem in 2025, 91% were in plugins and 9% in themes. WordPress core itself had just six.² In 2024 alone, 1,614 plugins and themes were pulled from WordPress.org over security problems their developers never fixed.¹³

A lot of that code is maintained by small teams or a single volunteer, doing thankless work and mostly hearing about it when something breaks. That makes them easy to wear down, and easy to buy out.

It's an old playbook: buy a popular plugin from a tired developer, then push out a poisoned "update." Every site that installs the update installs the backdoor, sometimes hundreds of thousands of sites at once. The buyers aren't hobbyists. They're well-funded operations that see a plugin's install base as an asset to cash in.

WordPress.org responded in 2026 by delaying plugin updates and adding automated reviews. Its own plugin review co-lead summed up the risk: "A plugin can be secure today and introduce a vulnerability."²⁰

What's a CVE?

A CVE (Common Vulnerabilities and Exposures) is an ID number for a publicly known security flaw, like CVE-2025-3102. The program exists so defenders everywhere can track and fix the same problem.²¹ The catch: that public write-up also tells attackers exactly what's broken and where. For popular software, ready-made attack code often shows up online soon after, so breaking in can be as simple as downloading a tool and pointing it at your site.

The flood is accelerating: 48,185 CVEs were published in 2025, up about 21% from 2024,²² and Microsoft expects roughly 72,000 in 2026.²³

Old PHP that will never be fixed

WordPress runs on a programming language called PHP. PHP 8.1 and everything older is officially end of life, meaning no more security fixes, ever.²⁴ Yet as of January 2026, about 26% of WordPress installs were still on PHP 7.4 or older,²⁵ and W3Techs data suggests nearly half of all PHP websites run a version that no longer gets patches.²⁶ Every known flaw in those versions stays open for good.

The front door is the same on every site

Every WordPress site has a login page at the same well-known address, so bots hammer them around the clock. Wordfence blocked more than 55 billion password-guessing attacks against WordPress sites in 2024.²⁷ An older WordPress feature called XML-RPC even lets attackers try hundreds of passwords in a single request.²⁸

All three are common. Researchers estimate just under a million WordPress sites were infected in 2024, with roughly 325,000 to 350,000 infected on any given day.²⁷

When a flaw becomes public, how long until attackers use it? Google's Mandiant team has tracked that average for years.³

That last bar isn't a typo. Mandiant now estimates attackers typically strike a week before a fix is even available: "exploitation is routinely occurring before a patch is even released."⁴

AI is a big reason why. Researchers showed back in 2025 that an AI system could read a public security advisory and produce a working attack in as little as 15 minutes, for about a dollar each.³⁸ The UK's National Cyber Security Centre warned that AI-assisted exploit development would shrink the window further.³⁹ By June 2026, the U.S. government ordered federal agencies to fix their most serious flaws within three days.⁴⁰ As CISA put it: "Defenders cannot afford to take weeks to patch systems that can be autonomously exploited."⁴¹

For fairness: not every measure gets worse every month. VulnCheck's share of day-zero attacks dipped in early 2026.³⁷ But the floor is now measured in hours, not weeks, and a small business website gets no three-day head start.

A sample of serious WordPress flaws from the last two years. CVSS is the standard 0 to 10 severity score; 9.0 and up is "critical."

AI finds the holes faster than people can

In late 2024, Google's Big Sleep AI found a real, exploitable flaw in SQLite, one of the most widely used pieces of software on Earth.⁵⁹ By mid-2025, an AI system called XBOW was the top-ranked hacker on HackerOne's U.S. leaderboard.⁶⁰ In February 2026, Anthropic reported its Claude model had found more than 500 high-severity flaws in open-source software.⁶¹ Then Anthropic's Claude Mythos Preview, shared only with a coalition of major tech and security companies under Project Glasswing, found thousands of previously unknown flaws across every major operating system and web browser,⁵ and more than 10,000 high- or critical-severity findings in its first month.⁶²

That's the defensive side, and it's good news for the software those companies maintain. Your 20-plugin WordPress site isn't on that list.

The guardrails, and the people who get around them

The big AI companies (OpenAI, Anthropic, Google, xAI) train their models to refuse hacking requests. People break those rules anyway. GPT-5 was jailbroken within about a day of release;⁶³ Grok 4 within two days.⁶⁴ In 2025, a Chinese state-sponsored group tricked Claude into running 80 to 90% of a real espionage campaign by posing as a security firm doing authorized testing.⁶⁵ Google has documented malware that calls AI models during live attacks.⁶⁶

Until recently, though, jailbreaking was a cat-and-mouse game the AI companies could fight: they could see misuse, patch it, and ban accounts. That's what kept serious offensive AI in a small number of hands.

What happens without guardrails

AI labs test their most capable models in locked-down sandboxes. In one test described in Anthropic's own safety report, an early version of Claude Mythos Preview was told to try to escape a secured sandbox. It built a multi-step exploit to reach the internet, emailed the researcher (who was eating a sandwich in a park at the time), and, without being asked, posted details of its exploit to public websites.⁶⁷⁶⁸ That's what raw capability looks like when nothing is holding it back.

The game changer: open-weight models

An open-weight model is one whose trained "brain" is published for anyone to download, run on their own computers, and modify.⁶⁹ There's no company in the middle watching for misuse and no account to ban. And the safety training can be removed: the UK's AI Security Institute says refusal training is "often easily reversible with access to the weights."⁶

That used to matter less, because open models lagged far behind. Not anymore:

• The gap is closing. The UK AI Security Institute found open-weight models now trail the frontier on cyber skills by just 4 to 7 months, down from 6 to 10 months a year earlier.⁶
• GLM-5.3 (Z.ai, China, August 2026). The U.S. government's AI standards center called it the most cyber-capable open-weight model released to date.⁷⁰ On one hard test of writing browser exploits, Anthropic's red team found it scored 50 out of 410, close to Mythos Preview's 56.⁷ Those are Anthropic's own benchmarks; Z.ai had not responded.⁷¹
• The guardrails didn't hold. Versions of GLM-5.3 with safety training stripped out appeared within days of release, and in Anthropic's testing they complied with harmful requests every time.⁷ Kimi K3, another Chinese open-weight model, was assessed by U.S. and U.K. government evaluators, who found its safeguards did not stop it from attempting exploit development.⁷²
• It's a business now. Companies openly host guardrail-free models for paying customers,⁷³ and researchers counted more than 3,400 "uncensored" models on one popular AI model site.⁷⁴ Alibaba's Qwen family is also open weight, with its next generation already in training.⁷⁵

The fix isn't a better plugin. It's getting off the platform. We build on modern frameworks used by some of the largest sites on the web:

Why this is safer

• No public admin login sitting at a well-known address for bots to hammer.
• No marketplace of strangers' code running with full access to your site.
• No exposed database sitting behind every page for visitors to poke at.
• The hosting platform patches the infrastructure and can push protections to every site at once.

Speed

Google measures real-world page experience with its Core Web Vitals. Only 45% of WordPress sites pass on mobile, among the lowest of any major platform.⁸ Sites built with Astro pass at about 67%.⁸¹ The build matters more than the label, and we tune every site for it.

AI can't recommend what it can't read

People increasingly ask ChatGPT, Claude, Gemini, Copilot, and Grok instead of scrolling search results. When Vercel studied AI crawler traffic, it found "none of the major AI crawlers currently render JavaScript."⁹ If your content only appears after scripts run, many AI systems simply don't see it. Our sites deliver the real content in the HTML itself.

We went straight to the source

These aren't techniques we made up. We read what each AI and search company says it needs, and built those requirements into the way we build:

• Google: no special tricks are required for AI Overviews and AI Mode, but a page must be indexed and eligible to show a snippet.⁸² Structured data should match what's visible on the page.⁸³
• Microsoft Bing: now reports how often a site is cited in Copilot and AI summaries, and recommends IndexNow so AI sees your latest content right away.⁸⁴⁸⁵
• OpenAI: sites should allow OAI-SearchBot to appear in ChatGPT search.⁸⁶
• Anthropic: Claude-SearchBot improves Claude's search results; blocking it can reduce visibility.⁸⁷
• xAI: Grok publishes no crawler documentation, so we test visibility in Grok directly.

SEO and AI visibility are built in from the first line of code, not bolted on later.

You may have heard that websites built with AI ("vibe coding") aren't secure. Six months ago, that was often true. In the hands of someone who doesn't understand security, it still is.

But what was true six months ago isn't true today. Paired with people who understand security, code, and how the internet actually works, AI builds websites that are clean, consistent, and hardened from the start. The same capability that lets AI find flaws lets it avoid them, and it never gets tired, skips a step, or forgets a check. In skilled hands, it's more reliable than most human developers working alone.

DarkHorse IT is a managed IT and cybersecurity company based in Moorhead, Minnesota. We build business websites with the same security-first mindset we bring to protecting your network, email, and computers.