Monday 30 March 2026Afternoon Edition

ZOTPAPER

News without the noise


Cybersecurity

OWASP Publishes Top 10 Security Risks for Model Context Protocol as CVEs Pile Up

Over 30 vulnerabilities filed in two months including shell injection and authentication bypass in widely used packages

Zotpaper2 min read
OWASP has published its first Top 10 security risks for the Model Context Protocol, the increasingly popular standard for connecting AI agents to external tools and data sources. The move follows a surge of over 30 CVEs filed against MCP servers, clients, and infrastructure between January and February 2026.

The Numbers

The breakdown of vulnerabilities is sobering. Forty-three per cent were exec or shell injection flaws where MCP servers passed user input directly to shell commands. Twenty per cent targeted tooling infrastructure including MCP clients, inspectors, and proxy tools. Thirteen per cent were authentication bypass issues where servers had no auth at all or implemented it incorrectly.

Critical Vulnerabilities

One vulnerability in a package downloaded nearly half a million times scored CVSS 9.6 for remote code execution. CVE-2026-27896 hit the official MCP Go SDK. Another high-severity bug targeted the Azure MCP server with an SSRF attack enabling privilege escalation.

What the Top 10 Covers

The OWASP MCP Top 10 is not a prediction about what might go wrong — it is a catalogue of what is already happening. The taxonomy covers the most common and impactful security failures in MCP deployments, from missing input validation to absent authentication to unsafe tool execution patterns.

Analysis

Why This Matters

MCP adoption has exploded in 2026 as AI agents become mainstream. Every major AI company now supports or builds on the protocol. But the security story has lagged far behind the adoption curve. Basic failures like shell injection and missing auth should not exist in production systems, yet they are the most common vulnerabilities.

Background

When OWASP publishes a Top 10 for something, it signals the security community has decided there is enough real-world risk to warrant a formal taxonomy. They did it for web applications, then for LLMs, and now for MCP.

Key Perspectives

The security community sees this as a wake-up call. MCP servers are effectively giving AI agents the ability to execute arbitrary actions on behalf of users, making security failures in this layer far more dangerous than traditional web vulnerabilities.

What to Watch

Whether MCP framework maintainers respond with mandatory security defaults. How quickly enterprises audit their existing MCP deployments. Whether this slows MCP adoption or simply matures it.

Sources