SystemGuard | Local Prompt Injection Scanner & Hardener

1. Load System Prompt

Paste LLM System Instructions

2. Adversarial Simulator Config

Adversarial LLM Model

WebGPU not supported. You can still use the static checklist and semantic audit features below.

Defense Score: {{ staticDefenseScore }}/100

Awaiting Prompt Evaluation

Load or type your prompt on the left, then click "Audit Prompt Defenses" to run structural tests.

Structural Auditing

{{ rule.name }}

Semantic Threat Analysis

We mapped your instruction phrasing against standard override vectors. Proximity of your structures to injection vectors:

Semantic Match Score: {{ (semanticLeakRating * 100).toFixed(0) }}%

Active Simulator Attack Sandbox

Run local overrides against instructions

Select Attack Payload

Simulated Completion Output {{ simulationVulnerable ? 'Injection Successful (Breached)' : 'Instruction Contained (Safe)' }}

Executing Local Model Inference...

Defensive Wrapping Frame

Encapsulates prompts inside containment blocks to neutralize injection inputs.

{{ hardenedPromptText }}

Understanding System Instruction Defenses

LLMs process system directives and user prompts inside the same unified context window. This architecture allows adversarial user inputs to easily override prior system instructions if no delimiters or isolation wrappers are used.

1. Delimiter Isolation

Enclosing system instructions inside stark dividers (such as triple hash signs ###) separates your code from untrusted input text. This acts as a clear signal boundary for parser layers.

2. Dynamic Red-Teaming

Automated tests using compact local models verify if instructions can survive override payloads without leaking underlying code blocks. This confirms prompt safety prior to production deployments.

3. Direct Leak Containment

Explicit instructions warning the model *never* to repeat, translate, or output initial system prompt files to end-users drastically reduce the risk of intellectual property exposure.