CVE-2026-125771

Read-only AI connections could change a WordPress site through prompts

AtlasAiDev · AtlasMCP (MCP Server for WordPress)

First documented disclosure · Review checkpoint

Summary

AtlasMCP allowed an AI app configured for read-only access to change a WordPress site through prompts.

Severity and exploitation evidence

High · 7.2/10 · CVSS 3.1. Severity source.

Exploitation not established: Patchstack describes the risk of exploitation. The reviewed advisory and changelog identify no confirmed malicious use.

Exploitation evidence source

Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.

How the weakness works

The prompt interface did not consistently enforce the connection’s read-only setting.

Conditions required

The maintainer describes a connected AI app with read-only access. Patchstack’s bulletin lists unauthenticated access and its CNA vector specifies high privileges, leaving the exact minimum privilege unresolved.

Potential impact

A connected AI client could make site changes beyond the access the WordPress operator had granted.

Affected versions and configurations

Through 2.0.3.

Fix and mitigation status

2.0.4.

Practical next steps

Upgrade MCP Server for WordPress to 2.0.4 or later. Review changes made by AI clients that were configured as read-only.

Operating lesson

Enforce connection permissions across prompts, resources and tools.

Evidence and limits

The November 11 maintainer changelog credits Patchstack and describes the read-only connection issue. Patchstack published its bulletin on November 13 and the CVE on November 15. Its earlier CVE-2026-121576 concerns a separate issue fixed in 2.0.2.

Sources