AI Governance Institute
← News
Research2026-11-17

87.5% of Known MCP Vulnerabilities Are Critical or High Severity

What happened

Researchers Xu He, Sagar Patel, and Shu Wang published Tracing MCP Security Vulnerabilities: Constructing and Analyzing an MCP Security Vulnerability Dataset at the 2026 IEEE Symposium on Security and Privacy Workshops. Starting from 119 MCP-related CVE records retrieved through February 2026, the team filtered out unrelated cases. This produced a validated snapshot of 24 vulnerabilities, with 96.7% agreement among reviewers on what to exclude. The Model Context Protocol (MCP) is the standard that lets AI agents connect to outside tools, databases, and services; it has become a primary interface in agentic AI systems. Of the 24 validated cases, 87.5% were rated Critical or High severity. Injection attacks combined with path traversal accounted for 66.7% of vulnerabilities. Path traversal is a technique where an attacker manipulates file paths to reach files or directories outside the intended scope. The researchers identified recurring patterns of insufficient input validation, weak access controls, and protocol-level gaps, and released a maintained dataset alongside reproducibility tooling for future research. This finding adds quantitative weight to a growing body of MCP security concerns. It follows prior reporting on 15,000 Public MCP Servers Found Vulnerable, Abandoned, or Sovereignty-Risky and 253 MCP Server Bugs and Six Framework Bypasses.

Why it matters

  • ·The 87.5% Critical or High severity rate means that most known MCP vulnerabilities carry the highest possible impact if exploited. Organizations using MCP servers as part of agentic AI deployments face a concentrated risk profile that general software patching programs may not be calibrated to handle.
  • ·The two dominant attack types are injection and path traversal. Injection smuggles malicious instructions into an AI agent's inputs. Path traversal reaches files outside permitted boundaries. Both are well-understood and defensible with specific controls. Their dominance at 66.7% of cases means targeted input validation and access restrictions, not broad security reviews, are the priority action.
  • ·Formal vulnerability management for MCP infrastructure is largely absent from enterprise AI governance programs. The dataset's snapshot covers only 24 cases from records through February 2026, and the maintained dataset is expected to grow. Organizations without an MCP server inventory or a process for tracking CVEs specific to agentic AI tooling have an unquantified and growing exposure.

Stay ahead of stories like this

Get every Global AI governance development like this one, plus the rest of the week's developments. Every Thursday.

Free every Thursday. Unsubscribe anytime.

Governance controls affected

What to do now

  • ☐Ask your engineering or IT team to produce a complete inventory of all MCP servers your organization runs or connects to, including those deployed by individual teams or developers outside central IT.
  • ☐For each MCP server in that inventory, confirm whether input validation controls are in place that reject malformed or unexpected commands before they reach the AI agent or connected tools.
  • ☐Verify that path restrictions are enforced on all MCP servers so that agents and tools cannot access files or directories outside their designated scope.
  • ☐Add MCP-specific CVE records to your existing vulnerability tracking process, using the publicly released dataset from this research as a baseline, and assign ownership for monitoring new disclosures.
  • ☐Include MCP server security controls in your next vendor due diligence review for any third-party agentic AI tools or platforms that use MCP to connect agents to external services.

What to watch next

The researchers released a maintained dataset extending beyond the original 24-record snapshot. The known MCP vulnerability count is expected to grow as the ecosystem matures and more disclosures are filed. Compliance teams should monitor emerging agentic AI regulations. This includes guidance under the EU AI Act (Regulation (EU) 2024/1689) and the Five Eyes Guidance on the Careful Adoption of Agentic AI Services. Teams should watch whether these reference MCP-specific controls as a named requirement. Regulators in financial services and critical infrastructure have signaled that agentic AI security is under active supervisory scrutiny. A growing CVE record for MCP infrastructure is likely to accelerate that focus.

Related Coverage

Standards2026-11-13

MCP and Agent Skills Move to Vendor-Neutral Governance Under Linux Foundation

Google reported in November 2026 that the Model Context Protocol (MCP) and Agent Skills specifications are being co-maintained under vendor-neutral governance through the Linux Foundation's Agentic AI Foundation. The transition means that the rules governing how AI agents connect to external tools and load procedural instructions are no longer controlled by any single vendor. Enterprise teams using MCP-connected agents must now track specification changes from a multi-party standards body and re-evaluate security controls at each revision.

Research2026-11-13

15,000 Public MCP Servers Found Vulnerable, Abandoned, or Sovereignty-Risky

OX Security analyzed over 15,000 MCP servers on public registries and 3,000 open-source AI projects, presenting findings at OWASP Global AppSec USA 2026. Researchers identified remote command execution risks, abandoned servers that attackers could take over, and data sovereignty concerns. Enterprises using AI agents that connect to MCP servers may be inheriting these risks without a systematic review process in place.

Standards2026-11-16

Microsoft Defense Guide Makes Layered Prompt Injection Controls an Enterprise Baseline

Microsoft has published updated guidance on defending against indirect prompt injection, a category of attack where malicious instructions hidden in external content redirect AI systems to perform unintended actions. The guidance, titled 'Defend against indirect prompt injection attacks,' recommends a layered set of controls including prompt shields, content isolation, behavioral monitoring, and human approval gates. It applies to any organization using AI systems that process third-party content such as emails, documents, websites, or plugins.