← Back to feed

Anthropic MCP Architectural Design Flaw: Zero-Click RCE Across 150M+ Downloads via Windsurf IDE

Date: 2026-04-25
Tags: malicious-tool, apt

Executive Summary

OX Security researchers uncovered an architectural RCE vulnerability in Anthropic's Model Context Protocol: Windsurf (CVE-2026-30615) was the only IDE where exploitation required zero user interaction, allowing remote attackers to execute arbitrary commands via prompt injection. The vulnerability ripples through 150M+ downloads and up to 200,000 vulnerable instances across Cursor, VS Code, Claude Code, and Gemini-CLI.

Campaign Summary

FieldDetail
Campaign / MalwareAnthropic MCP Architectural Design Vulnerability
AttributionVulnerability class; demonstrated by OX Security researchers (confidence: none)
TargetAI coding tools (Cursor, Windsurf, VS Code, Claude Code, Gemini-CLI); developers using MCP integrations
VectorSTDIO-based command execution via MCP configuration; zero-click prompt injection in Windsurf
Statusactive
First Observed2026-04-15

Detailed Findings

Anthropic's Model Context Protocol gives a direct configuration-to-command execution via their STDIO interface; in practice it actually lets anyone run any arbitrary OS command, if the command successfully creates an STDIO server it will return the handle, but when given a different command, it returns an error after the command is executed. Researchers identified four distinct families of exploitation: unauthenticated UI injection in popular AI frameworks; hardening bypasses in 'protected' environments like Flowise; zero-click prompt injection in leading AI IDEs (Windsurf, Cursor); and malicious marketplace distribution (9 out of 11 MCP registries were successfully 'poisoned' with a malicious trial balloon).

MITRE ATT&CK Mapping

TechniqueIDContext
Exploitation for Privilege EscalationT1548.004Arbitrary OS command execution via STDIO MCP configuration
Prompt InjectionT1597.002Zero-click indirect prompt injection in Windsurf IDE via HTML content
Supply Chain CompromiseT1195.001Malicious MCP packages distributed via 9/11 poisoned registries

IOCs

Domains

_Vulnerability is architectural, not isolated to single implementation; affects all Anthropic SDKs (Python, TypeScript, Java, Rust)_

Full URL Paths

_Vulnerability is architectural, not isolated to single implementation; affects all Anthropic SDKs (Python, TypeScript, Java, Rust)_

Splunk Format

_No IOCs available for Splunk query_

Detection Recommendations

Immediately audit all MCP STDIO configurations in .claude/settings.json and MCP marketplace repositories for suspicious commands; implement manifest-only execution or command allowlist in client applications (Cursor, Windsurf, VS Code); scan for MCP registries and plugins downloaded after April 15, 2026; monitor for 'config.patch' parameters and unauthenticated MCP connection attempts; deploy zero-trust execution model for all MCP tools with explicit user confirmation before any system command execution.

References