AI Guard

Behavioral multi-turn injection detection for LLM applications.

AI Guard models adversarial escalation across user turns instead of scanning a single prompt blob. It detects escalation drift, oscillation patterns, threshold gaming attempts, and computes a composite behavioral risk score (0–100).

Designed for:

Designed For

• LLM SaaS backends
• Prompt-based AI tools
• CI security enforcement
• Developers building AI workflows

---

Why AI Guard?

Most prompt injection tools scan a single prompt blob.

AI Guard models:

• User-turn escalation behavior
• Drift progression over time
• Oscillation patterns
• Late detection risk
• Threshold gaming attempts
• Composite adversarial risk scoring

It is behavioral, not just pattern-based.

---

Core Capabilities

• Transcript-aware analysis
• User-turn behavioral modeling
• Drift detection
• Oscillation detection
• Threshold gaming detection
• Composite behavioral risk score (0–100)
• CI threshold enforcement

---

Installation

Build from Source

git clone https://github.com/CODESHUB-O/ai-guard.git
cd ai-guard
cargo build --release

Binary will be located at:

target/release/ai_guard

---

Quick Start

Analyze a transcript:

./target/release/ai_guard multiturn examples/sample.txt --threshold 95

If the composite risk score is greater than or equal to the threshold,
the process exits with code 1 (CI failure).

---

Example Output

AI Guard — Behavioral Risk Analysis
------------------------------------
User Turns           : 4
Peak Severity        : Critical
Final Severity       : Critical
Drift Detected       : true
Partial Suppression  : false
Oscillation Detected : false
Late Detection       : false
Threshold Gaming     : false
Composite Risk Score : 72
Risk Level           : High
CI Threshold         : 95

---

JSON Output (CI Friendly)

./target/release/ai_guard multiturn transcript.txt --threshold 95 --json

Example:

{
  "composite_risk": 72,
  "risk_level": "High"
}

---

GitHub Action Example

name: AI Guard Scan

on:
  pull_request:
  push:
    branches:
      - main

jobs:
  ai-guard:
    runs-on: ubuntu-latest

    steps:
      - uses: actions/checkout@v4
      - uses: dtolnay/rust-toolchain@stable
      - run: cargo build --release
      - run: ./target/release/ai_guard multiturn examples/sample.txt --threshold 95

This will fail CI if behavioral risk exceeds the defined threshold.

---

Composite Risk Model (0–100)

Risk score is calculated using:

• Peak severity weight
• Detection delay
• Escalation velocity
• Drift presence
• Suppression patterns
• Oscillation detection
• Threshold gaming behavior

This enables behavioral adversarial modeling rather than static pattern checks.

---

Use Cases

• Secure LLM SaaS backends
• Validate prompt pipelines in CI
• Detect adversarial multi-turn escalation
• Monitor injection risk during development
• Guard AI tool execution workflows

---

Roadmap

Upcoming Pro Features:

• Repository-wide batch scanning
• Risk trend comparison across commits
• Historical baseline tracking
• Advanced reporting export
• Custom scoring configuration

---

License

MIT License

---

Philosophy

AI Guard focuses on behavioral modeling rather than surface-level string matching.

The goal is to detect adversarial progression patterns — not just isolated keywords.

Security is not static. Behavior matters.