Neural Prompt Injection Guard

Neural Prompt Injection Guard 🛡️

While regex and static heuristics intercept known exploit signatures in < 1 ms, sophisticated indirect prompt injections and jailbreak attacks require semantic evaluation. FaradAiWire provides an extensible Neural Semantic Security Guard (ADR-0032 & ADR-0042) powered by local Small Language Models (SLMs).

graph LR
    ToolCall["Tool Call Arguments"] --> FastHeuristics["Static Heuristics (<1ms)"]
    FastHeuristics -->|"Ambiguous / Untrusted"| Probe["Dual SLM Probe Engine"]
    Probe -->|"REST / OpenAI API"| Sidecar["Docker Sidecar (:8088)\n(Prompt-Guard)"]
    Probe -->|"Ollama API"| Ollama["Local Ollama (:11434)\n(Llama-Guard-3:1b)"]
    Probe -->|"Unix Socket / IPC"| MLX["Local Apple Silicon MLX\n(Qwen3Guard-0.6B)"]

    Sidecar --> Score{"Score >= 0.85?"}
    Ollama --> Score
    MLX --> Score
    Score -->|"Yes"| Block["Block Call (-32004)"]
    Score -->|"No"| Allow["Allow Upstream"]

Supported SLM Models & Engines

FaradAiWire interfaces with local inference engines running on developer workstations with zero cloud API keys:

Model ID	Architecture	Typical Latency	RAM Footprint	Provider
`prompt-guard`	Meta PromptGuard (86M)	12–18 ms	~250 MB	Docker Sidecar (`ghcr.io/faradailab/prompt-guard:latest`)
`llama-guard3:1b`	Llama Guard 3 (1B Q4)	22–35 ms	~1.1 GB	Local Ollama
`qwen3guard-0.6b`	Qwen 2.5 Guard (0.6B)	15–25 ms	~600 MB	Ollama / MLX / vLLM
`granite-guardian`	IBM Granite 3.0 (2B)	30–45 ms	~1.8 GB	Ollama

Quick Setup: Starting a Local Guard Engine

Option A: Local Ollama (Recommended for macOS & Linux)

Install Ollama.
Pull the lightweight guard model:
Terminal window
```
ollama pull llama-guard3:1b
```
FaradAiWire will automatically discover and link it on startup.

Option B: Docker Sidecar (Recommended for Containerized Environments)

docker run -d -p 8088:8088 --name prompt-guard ghcr.io/faradailab/prompt-guard:latest

Verifying & Managing Guards via CLI

FaradAiWire provides dedicated subcommands to inspect and manage inference engines:

Diagnose Providers:

faradaiwire doctor

Discovers running Ollama or Docker endpoints and tests round-trip inference latency.

List Installed Models:

faradaiwire guard list

AVAILABLE SLM MODELS:
 ID                   NAME                 PROVIDER    PARADIGM        STATUS
 prompt-guard-docker  Meta PromptGuard     docker      Classification  READY (8088)
 llama-guard3:1b      Llama Guard 3 1B     ollama      DecisionModel   READY (11434)
 qwen3guard-0.6b      Qwen3 Guard 0.6B     ollama      DecisionModel   AVAILABLE

Check Active Guard Status:

faradaiwire guard status

Configuration

In config.yaml:

filters:
  guard:
    enabled: true
    discovery:
      mode: "auto"              # "auto" (dual probe), "ollama", "docker", or "custom"
      ollama_url: "http://127.0.0.1:11434"
      docker_url: "http://127.0.0.1:8088"
      custom_url: ""            # Optional OpenAI-compatible endpoint (e.g. vLLM)
    confidence_threshold: 0.85  # Confidence score (0.00 to 1.00) required to trigger block
    timeout_ms: 1500            # Fallback timeout before failing closed or passing to fast-path

Visualizing Guard Decisions in the Web Console

Open http://localhost:8095/admin/dashboard and click the Guards tab:

Interactive Playground: Test arbitrary prompt injection payloads against your active local SLM to observe confidence scores and latency breakdowns.
Hardware Telemetry Bar: Real-time host CPU, RAM, and GPU memory usage.
GQA KV Cache Calculator: Accurately estimates memory requirements when scaling context windows from 2K to 8K tokens.