Preprint reports fewer successful attacks on AI agents with tool-call safeguards

A thesis posted on arXiv reports that two safeguards reduced average attack success against AI agents from 53.7% to 12.4% on MCPBench, a benchmark containing 847 scenarios. The safeguards target systems using the Model Context Protocol, which connects agents to tools. The proposed AttestMCP method authenticates tool calls with HMAC-protected packets at under 0.1 ms per call. It is paired with an isolation pattern called Commit Boundary. The methods are implemented in the MCPSec module. The reported reduction applies to benchmark scenarios, with successful attacks still occurring after the safeguards were applied.