Security Advisory

MCP Security Risks

Model Context Protocol enables powerful AI integrations, but introduces significant security risks. Here's what developers need to know.

Check if your app has vulnerabilities from AI-generated code.

What is MCP?

Model Context Protocol (MCP) is a standard created by Anthropic for connecting AI assistants to external tools. MCP servers run on your machine and give AI assistants like Cursor access to databases, APIs, file systems, and other resources. This power is what makes MCP useful - and also what makes it risky.

Security Risks

Arbitrary Code Execution

critical

MCP servers can execute arbitrary code on your machine with user-level privileges. This is by design - it's how MCP provides powerful functionality. However, malicious servers can abuse this to compromise your system.

Potential Impact

  • Full access to local files and directories
  • Ability to install malware or backdoors
  • Credential theft from environment variables
  • Network access to internal resources

Mitigation

Only install MCP servers from trusted sources. Review server code before installation.

Prompt Injection Attacks

high

Malicious content in files or external data can inject prompts that manipulate MCP server behavior. Attackers can craft inputs that cause MCP servers to execute unintended actions.

Potential Impact

  • Data exfiltration through manipulated requests
  • Unauthorized file modifications
  • Credential extraction via crafted prompts
  • Privilege escalation within the MCP context

Mitigation

Be cautious with MCP servers that process untrusted content. Validate inputs before processing.

Data Exfiltration

high

MCP servers have network access and can send data to external servers. Malicious or compromised servers could exfiltrate sensitive code, credentials, or personal information.

Potential Impact

  • Source code theft
  • API key and credential exposure
  • Personal information leakage
  • Corporate data breach

Mitigation

Monitor network activity from MCP processes. Use firewalls to restrict outbound connections.

Supply Chain Attacks

high

MCP server dependencies may be compromised. A popular MCP server could be updated with malicious code, affecting all users who update.

Potential Impact

  • Widespread compromise of developer machines
  • Backdoor installation through trusted updates
  • Credential harvesting at scale
  • Lateral movement into enterprise networks

Mitigation

Pin MCP server versions. Review changelogs before updating. Use servers from reputable maintainers.

Privilege Escalation

medium

MCP servers run with the same permissions as your IDE. If your IDE runs with elevated permissions or has access to sensitive resources, MCP servers inherit that access.

Potential Impact

  • Access to resources beyond intended scope
  • Modification of system configurations
  • Access to other users' data on shared systems

Mitigation

Run IDEs with minimal permissions. Use containerization for sensitive development work.

Known MCP Vulnerabilities

CVE-2025-54135

CurXecute - Slack to RCE via MCP

Critical

Prompt injection through Slack messages could trigger arbitrary code execution via Cursor MCP

CVE-2025-54136

MCPoison - Persistent Team Compromise

High

Malicious MCP server configuration could persist and spread across team environments

Safety Guidelines

Before Installing

  • Only use MCP servers from verified, reputable sources
  • Review the server's source code before installation
  • Check for recent security audits or reviews
  • Understand what permissions the server requires
  • Look for community feedback and issue reports

During Use

  • Monitor system resource usage by MCP processes
  • Watch for unexpected network connections
  • Be cautious when processing untrusted files or content
  • Review MCP server actions before approving
  • Keep servers updated, but review changes first

Environment Protection

  • Run development environment with minimal permissions
  • Use separate environments for sensitive projects
  • Consider containerization or VMs for isolation
  • Configure firewall rules for MCP processes
  • Keep sensitive credentials out of MCP-accessible paths

Affected Tools

ToolMCP SupportSecurity Notes
CursorNative MCP supportMCP servers can access all workspace files and execute terminal commands
Claude DesktopNative MCP supportMCP servers run with user permissions, can access configured resources
VS Code (with extensions)Via extensionsVarious extensions add MCP-like capabilities with varying security models
Custom MCP clientsVariesSecurity depends on implementation - some may have fewer restrictions

Check Your App's Security

Building with MCP-enabled tools? Our scanner checks for vulnerabilities commonly introduced by AI-generated code.

Scan Your App Free

Frequently Asked Questions

What is MCP?

MCP (Model Context Protocol) is a standard for connecting AI assistants to external tools and data sources. It allows AI coding assistants like Cursor to interact with databases, APIs, and local tools through 'MCP servers' that run on your machine.

Is MCP inherently insecure?

MCP is powerful by design - it's meant to let AI access tools and data. This power creates security risks if misused. The protocol itself isn't flawed, but its capabilities can be abused by malicious servers or through prompt injection.

Should I disable MCP?

You don't need to disable MCP entirely. Instead, only install MCP servers from trusted sources, review their code, and monitor their behavior. For sensitive work, consider using environments without MCP or with strict network restrictions.

How do I know if an MCP server is safe?

Check the source: Is it from a reputable organization or developer? Review the code: Does it do what it claims? Check for issues: Are there reported vulnerabilities? When in doubt, don't install it.

Can MCP servers access my entire system?

MCP servers run with your user permissions, so they can access anything you can access. They're typically scoped to specific resources by configuration, but a malicious server could ignore these boundaries.