# AtlasMCP: a read-only connection could still change a WordPress site

> A second escalation flaw in AtlasMCP let a WordPress connection marked Read only write to the site through prompts.

- **Severity:** Medium: Real and reproducible, but needs unusual setup, user help or a narrow audience. Plan a fix.
- **Category:** Vulnerability
- **Published by:** Patchstack
- **Disclosed:** 2026-11-13
- **CVE:** CVE-2026-125771
- **CVSS:** 7.2
- **Affects:** AtlasMCP, WordPress
- **Research by:** iml1s
- **Primary source:** https://patchstack.com/database/wordpress/plugin/ai-workflow-automation-ai-agent-hub/vulnerability/wordpress-mcp-server-for-wordpress-plugin-2-0-3-privilege-escalation-vulnerability
- **Page:** https://theinjection.dev/items/atlasmcp-read-only-connection-writes/

AtlasMCP, the WordPress plugin that connects Claude, ChatGPT and Gemini over MCP, let a connection set to Read only change the site through prompts in versions up to 2.0.3. Patchstack scores it 7.2 and the fix is 2.0.4.

## What happened

Patchstack published CVE-2026-125771 for the AtlasMCP WordPress plugin on November 13, 2026, ten days after a different escalation flaw in the same plugin was fixed. Patchstack credits iml1s, who reported the issue on November 10, 2026. AtlasAiDev shipped the fix in version 2.0.4 on November 11.

## How it works

AtlasMCP exposes a WordPress site to Claude, ChatGPT and Gemini as an MCP server, and its read-only setting covered the tools a connection could call but not its prompts. A connection marked Read only could therefore still change the site through the prompt surface. Version 2.0.4 makes prompts read-only, offers only prompts that match the tools a connection is given, and makes pausing AI access pause prompts and resources rather than tools alone.

## Who is affected

AtlasMCP 2.0.3 and earlier are affected, on a plugin that wordpress.org lists at 300 or more active installations. Sites that handed an AI app a connection labelled Read only are the ones that lost a guarantee they thought they had. The GitHub Advisory Database record scores the flaw with high privileges required, so the caller needs an existing connection to the site.

## What to do

Update AtlasMCP to 2.0.4 or later. Patchstack says it has issued a mitigation rule that blocks attacks on sites that cannot update yet. Sites that keep AI access paused should note that pausing only covers prompts and resources from 2.0.4 onward.

## Quick facts

- **CVE:** CVE-2026-125771
- **CVSS:** 7.2, from Patchstack and the GitHub Advisory Database
- **Affected:** AtlasMCP 2.0.3 and earlier
- **Fixed in:** 2.0.4, released November 11, 2026
- **Found by:** iml1s, reported through Patchstack on November 10, 2026
- **Earlier flaw:** CVE-2026-121576, a subscriber role change, fixed in 2.0.2 on November 3, 2026
- **Active installs:** 300-plus, per wordpress.org

## Questions

### Which AtlasMCP version fixes CVE-2026-125771?

AtlasMCP 2.0.4, released November 11, 2026, fixes CVE-2026-125771. Versions 2.0.3 and earlier are affected, and that includes 2.0.2, the release that closed the separate subscriber escalation flaw on November 3. Patchstack lists a mitigation rule for sites that cannot update immediately, and the vendor is AtlasAiDev.

### Is CVE-2026-125771 the same flaw as the AtlasMCP subscriber escalation?

No. CVE-2026-121576 let a signed-in subscriber change their own role through an AI app, and version 2.0.2 fixed that. CVE-2026-125771 is a different gap: the Read only setting on a connection governed tools but not prompts, so a read-only connection could still change the site. Different researchers reported the two issues through Patchstack.

### Does pausing AI access protect an unpatched AtlasMCP site?

Pausing is only a full stop from 2.0.4. The AtlasMCP changelog says that version made pausing AI access also pause prompts and resources, where earlier releases paused tools alone and left the prompt surface this flaw used. Patchstack's mitigation rule is the stopgap it names for sites that cannot update yet.

## Sources

1. [Patchstack Advisory](https://patchstack.com/database/wordpress/plugin/ai-workflow-automation-ai-agent-hub/vulnerability/wordpress-mcp-server-for-wordpress-plugin-2-0-3-privilege-escalation-vulnerability)
2. [GitHub Advisory](https://github.com/advisories/GHSA-w68m-3rvm-5xxq)
3. [Plugin Page](https://wordpress.org/plugins/ai-workflow-automation-ai-agent-hub/)
