LightGuard-Agent

🔍 Attack Scenario & Input

Preset Test Vectors:

Untrusted External Tool Observation / Payload: Proposed Tool Action (Requested by LLM):

📊 Defense Evaluation & Telemetry

Defense Latency

0.000 ms

Tier 2 Risk Score

0.00 / 1.00

Speedup vs Llama-Guard 3 (800 ms) 13,500x Faster

LightGuard executes in 59 µs with zero LLM generation overhead.

Tier 0: Steganography & Obfuscation Shield

Strips zero-width tokens, normalizes homoglyphs & leetspeak, decodes Base64/Hex.

Tier 1: Syntactic Sanitization

Neutralizes delimiters and redacts explicit injection directives.

Tier 2: Semantic Intent Classifier

Sub-0.03ms token heuristic density analysis.

Tier 3: Action Policy Firewall

Enforces execution policies and prevents unauthorized tool calls.

📈 Standardized Empirical Evaluation (50 Scenarios)

Benchmarked across 25 diverse Indirect Prompt Injection vectors and 25 benign control scenarios:

Defense Architecture	Attack Success Rate (ASR) ↓	False Positive Rate (FPR) ↓	Mean Latency (ms) ↓	Defense Strategy
Vanilla ReAct Agent	100.0%	0.0%	0.00 ms	None (Unprotected)
Regex-Only Filter	28.0%	0.0%	0.036 ms	Deterministic signatures only
LLM-as-a-Judge (Llama-Guard 3)	8.0%	4.0%	450 – 1200 ms	Heavy secondary model call
LightGuard-Agent (Ours)	0.0%	0.0%	0.059 ms	Tier 1 + Tier 2 + Tier 3 Pipeline