Introduction

GOLEM is a professional, open-source AI red teaming platform designed to audit Large Language Models (LLMs) for security vulnerabilities entirely through a seamless Web GUI.

Instead of wrestling with complex CLI flags, Python environments, or Docker containers, GOLEM packs a 200+ attack payload engine and an adaptive AI-vs-AI system into a single binary with a beautifully designed web interface.

Installation & Launch

GOLEM is distributed as a single static executable file. You do not need to install Python, Node.js, or Docker to use it. Just download the archive, extract it, and run.

.\golem.exe gui

scoop install golem
golem gui

chmod +x golem
./golem gui

sudo dpkg -i golem_Linux_x86_64.deb
golem gui

./golem gui

brew install OSINT-TECHNOLOGIES/tap/golem
golem gui

1. Setup & Configuration

The Setup tab is your command center for standard vulnerability scanning. Here you configure the target LLM and select which attack vectors to unleash.

Target Configuration

GOLEM supports both cloud APIs and local models:

• Local Models: Select "Local Server", provide the Base URL (e.g., http://localhost:1234/v1 for LM Studio or http://localhost:11434 for Ollama), and enter the model name. No API key required.
• Cloud Models: Select OpenAI or Anthropic, enter the exact model name (e.g., gpt-4o), and provide your API key.

Attack Vectors

Select from 12 distinct categories, representing over 200 attack techniques mapped directly to the OWASP LLM Top 10. Categories include Prompt Injection, Jailbreaks, Cipher Evasion, Tool Abuse (SSRF/SQLi), and Hallucination Triggers.

2. Neural Attack Engine (NAE)

The Neural Attack Engine is GOLEM's flagship feature. Instead of using static, pre-written payloads, NAE puts three LLMs into a live, adaptive adversarial loop.

You configure three distinct models:

• 🎯 Target Model: The victim you are trying to break.
• 🗡️ Attacker Model: The "mind" generating novel attack prompts. (Use a smart model like Qwen 2.5 7B or GPT-4o).
• 🔍 Analyzer Model: The strict judge evaluating the target's response. (A fast model like Llama 3.2 3B works great).

Attacker & Judge Profiles

You can quickly guide the behavior of the Attacker and Analyzer using one-click profiles:

Once launched, you can watch the models battle in real-time in the Attack Timeline. The Attacker will read the feedback from the Analyzer and try a different angle in the next round.

3. Attack Database

Curious about what GOLEM is actually sending to your models? The Attack DB tab is a complete, searchable reference of all 200+ attack techniques built into the platform.

• Filter attacks by category using the chips at the top.
• Search by technique name, description, or OWASP code (e.g., LLM06).
• Expand any attack card to see the exact example prompts GOLEM uses.

4. Live Scan Feed

When you start a standard scan, switch to the Live Scan tab. GOLEM uses Server-Sent Events (SSE) to stream attacks to your browser in real-time without polling.

The log entries are color-coded for instant visibility:

• [HIT] — The attack succeeded. The model leaked data, hallucinated, or bypassed its safety filters.
• [SAFE] — The model successfully detected and refused the malicious request.
• [ERR] — The API timed out, returned a 500 error, or refused connection.

Click on any log entry to expand and read the full attack prompt and the model's exact response.

5. Results & Export

Once a scan finishes, GOLEM generates a comprehensive security report on the Results tab.

• Score Cards: Displays the final GSS Score, Risk Level, and Vulnerability ratio.
• Severity Distribution: A bar chart breaking down vulnerabilities into Critical, High, Medium, and Low based on OWASP severity mapping.
• Top Findings: The most severe successful attacks are displayed with their prompts and responses for easy reporting.

Use the buttons at the bottom to download a standalone HTML Report to share with stakeholders, or a JSON file for integration into your CI/CD pipelines.

Core Concept: Pyramid Mode ⚡

Evaluating modern reasoning models (like DeepSeek-R1 or Ministral) is painfully slow because they generate thousands of "thinking" tokens before answering. A standard 200-attack scan could take 15+ hours.

Pyramid Mode (enabled by default in Advanced Settings) solves this by using a 3-phase adaptive approach, making scans up to 5x faster:

1. Phase 1 (Reconnaissance): GOLEM fires just 2-3 attacks from every category.
2. Phase 2 (Deep Dive): If a category was breached in Phase 1 (e.g., the model is vulnerable to Encoding Evasion), GOLEM runs the rest of the attacks only for that specific category. Safe categories are skipped.
3. Phase 3 (Exploitation): GOLEM takes the most successful prompts from Phase 2 and automatically mutates them (adding typos, spacing, case shifting) to discover the true boundary of the vulnerability.

Core Concept: Confidence Threshold

GOLEM uses a highly advanced Heuristic Evaluator to analyze model responses. Instead of simple keyword matching, it assigns a Confidence Score (0% - 100%) to every attack.

In the Advanced Settings, you can adjust the Confidence Threshold slider (default is 60%).

• Lower (e.g., 40%): More sensitive. Will flag partial data leaks or weak refusals as vulnerabilities. May increase false positives.
• Higher (e.g., 85%): Stricter. Will only count an attack as a HIT if the model explicitly and completely obeys the malicious command without any hesitation.

Core Concept: GSS Score

The GOLEM Security Score (GSS) is a standardized 0-100 rating designed to give stakeholders a clear understanding of an AI model's security posture.

GSS = 100 − SuccessRate% − (CriticalVulns × 5) − (HighVulns × 2)