GOLEM - Next-Gen security tester for AI
The AI Red Team in a single binary.
2 followers
The AI Red Team in a single binary.
2 followers
Audit any local or cloud LLM for jailbreaks, prompt injections, and data leaks in 60 seconds. GOLEM packs 200+ OWASP LLM attacks, a live AI-vs-AI Neural Attack Engine, and a dark Web GUI into one dependency-free Go binary. No Python. No Docker. 100% local.
This is the 2nd launch from GOLEM - Next-Gen security tester for AI. View more
GOLEM - Next-Gen security tester for AI
Launching today
Audit any local or cloud LLM for jailbreaks, prompt injections, and data leaks in 60 seconds. GOLEM packs 200+ OWASP LLM attacks, a live AI-vs-AI Neural Attack Engine, and a dark Web GUI into one dependency-free Go binary. No Python. No Docker. 100% local.
Free
Launch Team
Framer AI AgentsDesign and publish professional sites with AI
Promoted
Hey Product Hunt! 👋
We built GOLEM because testing LLMs for security vulnerabilities shouldn't take 3 hours of fighting broken Python virtual environments, dependency hell, and 40-page CLI docs.
Every week we see production AI apps leak their system prompts, fall for Base64/translation jailbreaks, or execute SQLi/SSRF via tool abuse. Existing red-teaming frameworks are great for academic research, but when you're shipping an AI feature or running a local model on Ollama/vLLM, you want answers in minutes, not hours.
### 🗿 Enter GOLEM: The AI Red Team in a Single Binary
Written in pure Go with zero external dependencies, GOLEM is a ~15MB executable. You download it, run `./golem gui`, and a real-time Web Dashboard opens on `localhost:8080`.
Here is what makes GOLEM different:
1. 🧠 Neural Attack Engine (AI vs. AI): Static jailbreak lists get patched in days. GOLEM puts three LLMs in an arena: an Attacker (which invents prompts), your Target model, and an Analyzer (judge). If your model refuses Round 1, the Attacker studies why it refused and adapts its strategy in Round 2 (e.g., switching to compliance framing + translation).
2. 🔺 Pyramid Mode (Built for Slow Reasoning Models): Auditing reasoning models (like DeepSeek-R1 or o1/o3) with 200+ prompts can take hours because of thinking tokens. Pyramid Mode runs a 3-phase smart scan: quick recon across all categories → deep dive only where the model flinched → automated payload mutation.
3. ⚔️ 200+ Attacks Mapped to OWASP LLM Top 10: Prompt injection, DAN/roleplay jailbreaks, system prompt extraction, homoglyphs, ciphers, crescendo chains, side-channel probing, and tool abuse.
4. 🔒 100% Local & Air-Gap Ready: No telemetry, no cloud accounts. Your API keys, prompts, and reports never leave your machine. Works out of the box with Ollama, LM Studio, vLLM, OpenAI, Anthropic, and any custom endpoint.
5. 📊 GOLEM Security Score (GSS): One-click standalone HTML reports for your team and JSON exports for CI/CD.
### ⚡ Try it in 60 seconds (100% Open Source & MIT Licensed)
We are launching `v0.1a` completely free and open-source today!
👉 GitHub Repo: https://github.com/OSINT-TECHNOLOGIES/Golem
👉 Website & Docs: https://osint-technologies.github.io/Golem/
How you can help us today:
If you're building with LLMs, run GOLEM against your system prompt or local model and tell us in the comments what GSS Grade (A–F) your model got!
And if GOLEM manages to break your model, dropping a ⭐ Star on our GitHub repository would mean the world to our team! 🙏