Researchers Discover 253 Security Vulnerabilities Across 343 AI Agent Tools in MCP Servers

Researchers published the ToFu study demonstrating 253 verified security vulnerabilities in 343 AI agent tools across 118 MCP servers; using a fuzzing framework that maps tool metadata to semantics-aware sanitizers, they found issues including path traversal, log injection, data exfiltration, and command bypass, and showed how model-generated arguments (via a "Repeat-After-Me" attack) can be misinterpreted as executable tool calls—the paper recommends enforcing documented runtime restrictions and treating schema-valid inputs as untrusted.