Breaking the Control Plane: Exploiting MCP Servers in AI Workflows

Co-Founder, CTO, Pluto Security

MCP servers are increasingly deployed in production environments to bridge agents with SaaS platforms, source control, CI/CD pipelines, and internal APIs. In practice, these services:

‍

  • Hold API tokens and personal access credentials
  • Perform outbound HTTP requests
  • Read and write to local filesystems
  • Execute privileged automation steps
  • Are frequently exposed beyond localhost, often bound to 0.0.0.0 by default

‍

Through a broad review of open-source MCP implementations and deployment patterns, we identified recurring security weaknesses across the ecosystem, including:

‍

  • Insecure network exposure defaults
  • Middleware-level trust boundary failures
  • Unvalidated outbound request destinations
  • Unconstrained filesystem access
  • The assumption that MCP servers are "local dev tools" rather than infrastructure components

‍

We will demonstrate practical exploitation paths that arise from these design patterns, including SSRF pivoting, arbitrary file writes, and escalation to remote code execution. We will also present multiple real-world example of vulnerabilities disclosed as part of the research in some of the most popular and widely used MCP servers, illustrating how classical issues compound in agent-driven architectures. Examples will include: CVE-2026-27826 & CVE-2026-27825, allowing a critical unauthenticated RCE chain in Atlassian MCP server (4M+ downloads), CVE-2026-61559 & CVE-2026-61560, allowing for full account takeover via the most popular Gitlab MCP, CVE-2026-33032 a critical NginxUI MCP vulnerability allowing any attacker on your network to take complete control over your nginx web servers, and more...

‍

In addition, we explore the growing ecosystem of online MCP server and agent "skill" marketplaces, and discuss the emerging supply chain risks introduced by composable AI integrations with minimal security review.

‍

Rather than focusing on model behavior or prompt injection, this talk examines the infrastructure layer enabling AI workflows and highlights why MCP servers should be threat-modeled and hardened as first-class production services.

‍

‍