Agentic Boundary & AST Firewall Lab

Enterprise Agentic Prompt Injection & Boundary Firewall Simulator

Simulate and benchmark deterministic AST grammar syntax firewalls vs secondary guardrail LLMs. Model sub-2ms zero-allocation boundary checks against indirect prompt injection, tool parameter SQLi, and token smuggling.

Multi-Tier Defense Architecture Benchmark

Elesium AST Syntax Firewall

Deterministic pre-execution AST parsing and Pydantic schema whitelist boundaries validate payloads in sub-2ms with zero token overhead and near-absolute 99.8% block rates.

Latency: 1.2ms | Block Rate: 99.8% | Security Cost: $0/mo

Secondary Guardrail LLM (Llama-Guard-3 / NeMo)

Secondary model evaluation adds 220-280ms latency penalties, duplicates 100% of token consumption, and suffers from 5.2% false positives on legitimate SQL and code.

Latency: 240ms | Block Rate: 84.0% | False Positive: 5.2%

Naive Direct LLM Ingestion

Zero input sanitization permits prompt injections, markdown image JWT leaks, and parameter SQLi, resulting in full agent tool compromise.

Latency: 0.0ms | Block Rate: 0.0% | Compromise Risk: Critical