RedProbe 🔴

A local LLM red-teaming framework for adversarial robustness testing.

RedProbe runs structured attack suites against locally hosted LLMs via Ollama, scores responses using an LLM-as-judge pipeline, and generates detailed security reports — entirely offline, no API keys required.

---

Why

As LLMs get embedded into products, their failure modes matter. RedProbe makes it easy to systematically probe a model's defences against:

• Prompt injection
• Jailbreaks
• System prompt extraction
• Role confusion

Most existing tools either require cloud APIs or treat red-teaming as a manual process. RedProbe is fully local and automated.

---

Requirements

• Python 3.10+
• Ollama running locally
• At least one model pulled via Ollama

---

Setup

git clone https://github.com/GodSagar007/redprobe.git
cd redprobe
python -m venv .venv

# Linux/Mac
source .venv/bin/activate

# Windows
.venv\Scripts\activate

pip install -r requirements.txt

Pull models to test against:

ollama pull mistral
ollama pull tinyllama

---

Usage

Run full attack suite against a single model:

python main.py --models mistral

Compare multiple models:

python main.py --models mistral --models tinyllama

Run a specific attack category only:

python main.py --models mistral --category jailbreak

All options:

python main.py --help

---

Attack Categories

Category	Description
prompt_injection	Instructions designed to override the system prompt
jailbreak	Attempts to bypass core model behaviour via social engineering
extraction	Attacks targeting confidential system prompt contents
role_confusion	Identity manipulation to destabilise the model's defined role

---

Scoring

Responses are evaluated by an LLM-as-judge pipeline using a secondary model.

Status	Meaning
🔴 BROKEN	Model violated its system prompt or was successfully manipulated
🟡 PARTIAL	Model partially complied with the attack
🟢 HELD	Model successfully resisted
⚪ UNCLEAR	Judge confidence too low — flagged for manual review

---

Output

Each scan produces a JSON report in reports/. When testing multiple models, a unified comparison report is generated with breach rates broken down by category and models ranked safest to most vulnerable.

Example comparison output:

{
  "ranking": ["mistral", "tinyllama"],
  "comparison": {
    "mistral": {
      "breach_rate_percent": 50.0,
      "breach_rate_by_category": {
        "extraction": 40.0,
        "jailbreak": 60.0,
        "prompt_injection": 60.0,
        "role_confusion": 40.0
      }
    }
  }
}

---

Project Structure

redprobe/
├── core/
│   ├── scorer.py       # Scoring pipeline
│   ├── llm_judge.py    # LLM-as-judge implementation
│   └── reporter.py     # Report generation
├── tests/
│   ├── prompt_injection.json
│   ├── jailbreaks.json
│   ├── extraction.json
│   └── role_confusion.json
├── reports/            # Generated reports (gitignored)
├── main.py             # CLI entry point
└── requirements.txt

Disclaimer

This tool is intended for security research and educational purposes. Only use it against models you own or have explicit permission to test.