AI MCP Servers in Cybersecurity: Emerging Attack Vectors and Mitigation Strategies
Year: 2025
Type: conference-paper
Abstract: As the rapid adoption of Model Context Protocol (MCP) servers in artificial intelligence (AI) ecosystems enables seamless agent-to-agent communication, tool integration, and enterprise-level orchestration, this accelerated adoption leads to a vast and expanding attack surface that has not been primarily explored in the cybersecurity literature. This manuscript explores the attack surfaces emerging as a direct result of AI MCP servers, including protocol-level attacks, the unauthorized invocation of tools, supply chain risks, data exfiltration, and denial-of-service risks. We also discuss how malicious or compromised tools can act as accelerators of an entire system’s vulnerabilities. To work against these new attack surfaces, we define a taxonomy of MCP-specific threats and describe a set of mitigation strategies based on principles of defense-in-depth and zero-trust. The mitigation strategies include: protocol hardening, authentication and authorization measures, runtime security monitoring, and supply chain integrity validation. We explore the risks and defenses of AI MCP servers in a systematic and structured way to provide a foundation for the next-generation AI systems, as discussed with both the researchers and practitioners of the AI field.
Authors: Thirumalaisamy, Karthikeyan, Konakalla, Manikarthik, Devamanoharan, Dinesh Kumar
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7115185181
FWCI:
Cites:
Cited by:
Related to: