Why MCP servers are becoming AI’s newest attack surface
ainews ·
MCP servers have rapidly become the standard way for AI agents to connect with external tools and data, spreading from Anthropic’s launch in November 2024 to more than 10,000 public servers by December 2025. Their adoption by major platforms and coding assistants has improved interoperability but created a fast-growing attack surface that existing AI defenses were not designed to handle. Threats include tool poisoning, rug pulls, tool shadowing, cross-origin escalation, data exfiltration, and excessive permissions. A Lakera analysis of 10,000 MCP servers found that 40% had exploitable weaknesses. The article distinguishes MCP security from broader agent security, noting that agents can access systems through channels other than MCP and therefore require additional controls. Vendors including TrueFoundry, Cisco, and Check Point are developing governance tools, guardrails, traffic inspection, and AI-focused firewalls to address these risks.
ainews