← All insights
Security & Trust

WordPress 7.0 Shipped an AI Agent Platform. Hackers Got the Keys on Day Two.

WordPress 7.0 'Armstrong' added a Connectors API that stores Anthropic, Google, and OpenAI keys in wp_options. Patchstack's founder called it 'free AI tokens for hackers.' AI scanning found 300+ zero-days at $20 each in 72 hours. SiteGround pushed 1M+ installs automatically.

· 8 min read
Share on X LinkedIn

The Feature That Became a Target

WordPress 7.0 'Armstrong,' released May 20, 2026, shipped the most ambitious update in WordPress history: a full AI agent infrastructure. WP AI Client provides a unified interface to Claude, Gemini, and GPT. The Connectors API stores API keys. The Abilities API exposes site operations to AI models. And a WordPress MCP Adapter lets AI coding agents manage WordPress installations directly.

Within two days, Patchstack founder Oliver Sild published the warning that the security community was already thinking: 'WordPress 7.0 combined with plugin vulnerabilities equals free AI tokens. There will be an absolute rush by hackers to steal API keys.' The keys sit in wp_options — the same database table that every SQL injection vulnerability in WordPress history has targeted.

2 days
Time from WP 7.0 launch to security warning
Patchstack founder Oliver Sild published his warning May 22, 2026. Source: Patchstack, Search Engine Journal.
3
API providers whose keys are stored in wp_options
Anthropic, Google, OpenAI via the Connectors API. Source: WordPress 7.0 documentation.

The $20 Zero-Day Assembly Line

Simultaneously, researchers demonstrated that AI-powered vulnerability scanning could find WordPress zero-days at industrial scale and negligible cost. 300+ critical vulnerabilities discovered in 72 hours. Average cost per zero-day: approximately $20 in compute time. Categories found: pre-authentication remote code execution, SQL injection hidden behind PHPCS annotations, privilege escalation through WordPress hooks, server-side request forgery, and downgrade attack chains.

The math is devastating. WordPress 7.0 stores high-value API keys in the same database accessible through the same vulnerability classes that AI scanners can now discover for $20 each. The attack surface expanded. The cost to exploit it collapsed. Both happened in the same month.

300+
Zero-days found by AI scanning in 72 hours
Source: Help Net Security, May 2026. Includes pre-auth RCE, SQLi, privilege escalation.
~$20
Average cost per AI-discovered zero-day
Source: Help Net Security. Compute cost for AI-powered vulnerability discovery.

The SiteGround Distribution Problem

SiteGround, one of the largest WordPress hosting providers, pushed its AI Agent plugin across its entire hosting network. Over 1 million installations — not through user choice, but through automatic distribution as part of hosting infrastructure. Site owners who never asked for AI agent capabilities received them. Their sites now store API keys, expose MCP endpoints, and present attack surfaces that didn't exist before the hosting provider's update.

This is the WordPress distribution model working as designed. The same automatic update mechanism that pushes security patches also pushes new attack surface. The same hosting provider trust that keeps sites patched also enrolls them in AI agent experiments.

1M+
SiteGround AI Agent auto-installations
Source: SiteGround. Distributed via hosting platform, not user opt-in.

The Convergence WebPulse Predicted

WebPulse has been tracking two independent crisis vectors: WordPress's security architecture (18,005 CVEs, 11,334 new vulnerabilities in 2025, 5-hour median exploitation window) and the rise of AI agents as the dominant web consumers (57.5% bot traffic, 1,000x page multiplication). WordPress 7.0 merged them into a single attack surface.

A WordPress site with the Connectors API enabled is simultaneously: a target for traditional web exploitation (SQL injection, plugin vulnerabilities), a source of high-value AI API keys (worth actual money on underground markets), and an MCP endpoint that AI agents can discover and interact with. Three attack surfaces on one CMS.

The frameworks that don't store API keys in databases — because they don't have databases — don't have this problem. Hugo generates static HTML. Astro generates static HTML. Neither stores secrets in a SQL-injectable table. In the AI agent era, the architectural decision that matters most is whether your framework stores valuable credentials in a place attackers already know how to reach.

Share this insight
More insights