Overview
Re-implemented and evaluated four security frameworks — SAMOS, AgentBound, MCPShield, and Breaking the Protocol — on a dataset of 299 MCP servers and 27,861 metadata entries. Controlled experiments covered least-privilege capability boundaries, gateway information-flow control, metadata misalignment detection (78.38% detection rate), and protocol-level vulnerabilities. Reproduced metrics were benchmarked against the original papers, highlighting limitations in LLM agent tool integration and defense-in-depth strategies.
Methods
- Controlled evaluation
- Information-flow control
- Capability restrictions
- Metadata misalignment detection