Overview & Adversary Mindset

OSAI is OffSec's offensive AI security certification (AI-300). It applies the same adversary-first methodology from OSCP to AI-enabled systems — LLMs, agents, RAG pipelines, ML infrastructure. This isn't about prompt engineering. It's about breaking AI systems like an attacker.

What OSAI Tests

• Identifying and exploiting vulnerabilities in LLM-backed applications
• Attacking Retrieval-Augmented Generation (RAG) pipelines
• Compromising multi-agent AI systems and their tool-call surfaces
• Exploiting AI deployment infrastructure (serving stacks, APIs, model files)
• Combining classic offensive techniques with AI-specific attack primitives

The AI Attacker's Mindset

Traditional pentesting assumes deterministic systems — same input, same output. AI breaks this. You're attacking a probabilistic system that changes behavior based on context, temperature, and sampling. This changes how you:

• Reproduce findings — outputs vary. Document prompts, not just outputs
• Validate exploits — run multiple times to confirm reliability
• Communicate risk — probabilistic failure needs statistical framing
• Iterate — failed attacks need reformulation, not abandonment

Think of an LLM as a software system with a natural language interface. Every input you give is a "function call". Your job is to find the unintended code paths — just like SQL injection, but for language models.

Prerequisites Checklist

• Linux CLI
• Python
• HTTP/APIs
• Basic LLM Concepts
• Offensive Fundamentals