pentest-ai-llm-security

# Pentest AI/LLM Security ## Purpose AI-integrated applications introduce entirely new attack surfaces. Prompt injection is the "SQLi of AI." Neither Shannon nor any existing skill addresses this domain. OWASP LLM Top 10 (2025) defines the methodology. ## Prerequisites ### Authorization Requirements - **Written authorization** with AI/LLM testing scope explicitly included - **Model access details** — API endpoints, model versions, tool/function access - **Data sensitivity classification** — what data the LLM can access - **Rate limit awareness** — LLM API costs can escalate quickly ### Environment Setup - Garak for automated LLM vulnerability scanning - Burp Suite for API interception of LLM requests/responses - Python scripts for custom prompt injection payloads - Local proxy to capture full request/response chains ## Core Workflow 1. **Integration Point Discovery**: Identify all LLM integration points — chat interfaces, content generation, RAG pipelines, AI search, code completion, summarization. 2. **Direct Prompt Injection**: Override system prompts, extract system prompt content, inject instructions that change model behavior. 3. **Indirect Prompt Injection**: Embed malicious instructions in documents/emails/web pages the LLM processes, poisoned RAG context. 4. **Data Exfiltration**: Extract training data, PII from context windows, other users' conversation history, system config details. 5. **Insecure Output Handling**: LLM output rendered as HTML (XSS via LLM), used in SQL queries (SQLi via LLM), used in system commands. 6. **Excessive Agency**: LLM with tool access performing unauthorized actions, privilege escalation through tool chains, resource abuse. 7. **Classification**: Document findings with OWASP LLM Top 10 (2025) classification and remediation guidance. ## OWASP LLM Top 10 (2025) Coverage | Category | Test Focus | Status | |----------|-----------|--------| | LLM01 Prompt Injection | Direct and indirect injection | ✅ | | LLM02 Sensitive Information Disclosure | Data exfiltration, PII leakage | ✅ | | LLM03 Supply Chain | Model provenance, plugin trust | ✅ | | LLM04 Data and Model Poisoning | Training data integrity | ✅ | | LLM05 Improper Output Handling | XSS/SQLi via LLM output | ✅ | | LLM06 Excessive Agency | Unauthorized tool use | ✅ | | LLM07 System Prompt Leakage | System prompt extraction | ✅ | | LLM08 Vector and Embedding Weaknesses | RAG poisoning | ✅ | | LLM09 Misinformation | Hallucination exploitation | ✅ | | LLM10 Unbounded Consumption | Resource exhaustion | ✅ | ## Tool Categories | Category | Tools | Purpose | |----------|-------|---------| | LLM Scanning | Garak, rebuff | Automated prompt injection testing | | API Interception | Burp Suite, mitmproxy | LLM API request/response capture | | Prompt Fuzzing | Custom Python scripts | Payload generation and testing | | Output Analysis | Browser DevTools, Burp | Insecure output rendering detection | ## References - `references/tools.md` - Tool function signatures and parameters - `references/workflows.md` - Attack pattern definitions and test vectors