Back to Glossary
PROTO

Deterministic Proxy

non-probabilistic filteringMCP proxy architecturerule-based enforcement

A deterministic proxy applies rule-based, non-probabilistic security enforcement to all communications between an AI agent and its tool ecosystem.

WHY PROBABILISTIC GUARDRAILS FAIL

LLM safety layers are trained on distributions. Adversarial inputs target the boundaries of these distributions, pushing the model beyond its guardrails. A deterministic proxy has no such boundary.

PROTOCOL CONTEXT (MCPVANGUARD ARCHITECTURE)

McpVanguard is implemented as a deterministic proxy between the AI orchestrator and tool server layer. Rules are loaded from versioned YAML manifests and evaluated in a high-throughput engine — not a language model.

ProvnAI Mitigation

McpVanguard's proxy is implemented as a low-latency Rust service. Versioned rule manifests are signed and recorded in the Merkle Audit Trail, ensuring the active rule set is auditable and tamper-evident.