analytics EMPIRICAL BENCHMARKS

442 Adversarial Attacks, Measured Locally

Every performance and security claim is backed by deterministic reproducible benchmarks. Run python run_comprehensive_benchmark.py to verify all 442 test vectors on your machine.

328 / 442

74.2% Pass Rate (Phase 1)

Local deterministic regex & AST heuristics alone with zero network latency.

350 / 442

79.2% Pass Rate (Phase 2)

Combined with optional Tier-2 Groq LLM-as-a-Judge semantic arbiter.

0.12 ms

p50 CPU Latency

Measured over 1,000 iterations on standard single developer CPU core.

100% Offline

Zero Network Calls

Runs completely inside process memory with zero external dependencies.

Empirical Test Results by Category

Evaluated against the comprehensive adversarial suite

Verified Deterministic

Attack Category	Test Cases	Blocked / Allowed	Resolution Rate	Defense Layer
Basic Prompt Injection Overrides	10	10 / 10	100% BLOCKED	Regex + Semantic Hierarchy
Instruction Hierarchy & Delimiters	10	10 / 10	100% BLOCKED	Pseudo-header & XML Tag Stripper
Unicode & Homoglyph Steganography	10	10 / 10	100% BLOCKED	NFKD + Cyrillic Bounded Normalizer
Educational Trojan Bypasses	10	10 / 10	100% BLOCKED	Trojan Framing Detector
Educational Discussion (Harmless)	10	10 / 10	100% ALLOWED	Referential Inquiry Whitelist
Destructive Shell Commands (rm -rf)	11	11 / 11	100% BLOCKED	CommandGuard Binary Whitelist
Destructive SQL DDL & Drops	10	10 / 10	100% BLOCKED	Literal-Aware SQL Comment Stripper
Output Structural JWT & API Keys	10	10 / 10	100% SCRUBBED	Base64url Structure Validator

Micro-Benchmark Latency Profile

Measured over 1,000 iterations per detector module on a single CPU core

PII & Luhn Validator

0.089 ms p50

Prompt Injection Guard

0.124 ms p50

Command & SQL Guard

0.080 ms p50

Secret & JWT Scrubbing

0.032 ms p50

Reproduce Benchmarks on Your Machine

Run the comprehensive benchmark script to re-evaluate all 442 test cases locally.

python run_comprehensive_benchmark.py