Skip to content

README.md

Latest commit

 

History

History
198 lines (133 loc) · 3.92 KB

README.md

File metadata and controls

198 lines (133 loc) · 3.92 KB

LATT

LLM Attack Testing Toolkit

🇺🇸 English | 🇪🇸 Español

A structured methodology and mindset framework for testing Large Language Model (LLM) applications against logic abuse, prompt injection, jailbreaks, and workflow manipulation.

What is LATT?

LATT (LLM Attack Testing Toolkit) is a practical, human-driven testing framework focused on identifying vulnerabilities in LLM-powered systems.

It is not a fuzzing tool. It is not an automated scanner.

It is a thinking toolkit.

LATT helps security researchers, red teamers, and builders systematically explore:

  • Prompt injection
  • Jailbreak techniques
  • Instruction override
  • Context leakage
  • Tool misuse
  • Workflow manipulation
  • Logic abuse in AI-assisted systems

Philosophy

LLM security is not only about payloads — it is about flows, trust boundaries, and hidden assumptions.

Traditional web testing looks at:

  • Input → Processing → Output

LLM testing must also consider:

  • System prompt
  • Tool calls
  • Memory/state
  • Role context
  • Multi-turn interaction logic
  • Human-in-the-loop assumptions

LATT is designed to map and break those assumptions.

Repository Structure

LATT/
│
├── COLLECTION.md  # Main Collection of pentesting elements
├── CYOE.txt       # GenAI OffSec Testing Guide based on LLM
├── LICENSE        # MIT License for this repository
├── README.md      # English documentation
└── README_ES.md   # Spanish documentation

Files

  • COLLECTION.md

    • Structured attack surface checklist
    • Categorized testing strategies
    • Reusable mental models
    • Payload patterns and abuse techniques

  • CYOE.txt

    • “Choose Your Own Exploit”
    • Interactive red-team simulation
    • Flow-based decision testing
    • Designed to train attacker intuition

What LATT Covers

1. Prompt Injection

  • Instruction override
  • Hidden directive hijacking
  • Context poisoning
  • Multi-layer prompt stacking

2. Jailbreak Techniques

  • Role confusion
  • Fictional framing
  • Encoding/obfuscation
  • Indirect instruction smuggling

3. Context & Memory Abuse

  • Session leakage
  • Cross-user contamination
  • Retrieval poisoning
  • State manipulation

4. Tool & Agent Abuse

  • Tool call manipulation
  • Argument injection
  • Over-permissioned integrations
  • External system pivoting

5. Workflow & Logic Abuse

  • Approval bypass
  • Policy misalignment
  • Safety guard chaining
  • Business logic exploitation

Intended Audience

  • Security researchers
  • Bug bounty hunters
  • Red team operators
  • AI product security engineers
  • Developers building LLM-powered applications

How To Use LATT

  1. Start with COLLECTION.md

    • Identify relevant attack categories
    • Build your testing hypotheses

  2. Use CYOE.txt

    • Simulate decision trees
    • Train adversarial thinking
    • Explore multi-step exploit paths

  3. Apply to real-world LLM systems:

    • Chatbots
    • AI copilots
    • Agent-based workflows
    • Retrieval-augmented generation (RAG) systems
    • AI integrations with APIs or internal tools

Design Principles

  • Human-driven exploration > blind automation
  • Context-aware testing
  • Flow-based attack modeling
  • Minimal noise, maximum signal
  • Offensive mindset with defensive value

Non-Goals

  • This is not a payload dump.
  • This is not a wordlist repository.
  • This is not an automated exploitation framework.

LATT focuses on thinking frameworks, not raw tooling.

Contributing

If you want to contribute:

  • Add structured attack patterns
  • Improve categorization clarity
  • Expand CYOE branches
  • Share anonymized case studies

Keep contributions structured and methodology-focused.

Disclaimer

This toolkit is provided for educational and defensive security research purposes only.

Use responsibly.

Made with <3 by URDev.