01 // What is C-AI/MLPen?
The Certified AI/ML Pentester (C-AI/MLPen) is a practical certification by The SecOps Group that tests your ability to find and exploit vulnerabilities in AI-powered applications. Unlike theory-heavy certifications, this one is entirely hands-on: you're given AI chatbots to break, and your job is to extract their secrets.
The exam covers the full spectrum of AI/LLM attack techniques: prompt injection, NL2SQL injection, indirect injection via file uploads, hallucination exploitation, and creative prompt engineering. Each challenge is a different AI chatbot with progressively stronger defenses - from basic input filters to multi-layered guardrails with output filtering, persona enforcement, and decoy outputs.
02 // Exam format
The exam consists of 8 practical challenges, each featuring an AI chatbot guarding a secret flag. You interact with the bots through a web interface and try to extract the hidden values using prompt injection techniques. Flags are submitted via the exam portal for instant validation.
| Detail | Value |
|---|---|
| Questions | 8 practical challenges |
| Format | CTF-style - interact with AI chatbots, extract flags |
| Time | ~3 hours recommended (no hard cutoff observed) |
| Passing score | Not publicly disclosed |
| Proctoring | None - VPN-based access to exam platform |
| Retakes | 1 free retake within 14 days |
| Validity | Lifetime |
03 // What skills do you need?
The exam tests practical prompt hacking skills. You need to understand how LLMs process instructions, where trust boundaries exist in AI applications, and how to creatively bypass filters and guardrails. No single technique works on every level - the exam is designed to force you to adapt your approach.
Traditional pentesting skills translate directly: thinking outside the box, persistence when stuck, and a systematic methodology are what separate someone who solves 3 levels from someone who solves all 8. If you've done web app pentesting or CTFs, the mindset is the same - just applied to a new attack surface.
04 // How to prepare
Study the OWASP Top 10 for LLMs (2025) and MITRE ATLAS. Understand how LLMs work at a conceptual level: tokens, prompts, system prompts, context windows, and inference. You don't need to train models, but you need to know how they process input.
Gandalf (Lakera AI), Prompt(air)lines, HackaPrompt, and myllmbank.com are the closest to the actual exam format. If you can beat Gandalf's hardest levels, you have the right skillset. Also check learnprompting.org for prompt hacking techniques.
Build a personal cheat sheet of techniques: direct/indirect injection, refusal suppression, context switching, obfuscation (base64, ROT13, token smuggling), payload splitting, virtualization, few-shot attacks, DAN-style jailbreaks, and dictionary attacks. Know when each one applies.
INE's AI Security (eAIS) learning path covers the theory behind all these attacks across 7 courses. Not required for the exam, but gives solid foundations if you're new to AI security. I studied it alongside and it helped with understanding the "why" behind each technique.
05 // My experience
The exam is well-designed and genuinely tests practical skills. Each level requires a different approach - there is no silver bullet prompt that works everywhere. The difficulty ramp is fair: early levels fall quickly, while later ones require creative thinking and persistence.
The most interesting challenges were the ones involving AI systems connected to real backends - databases, file systems, and document stores. These test whether you can exploit the full AI application stack, not just the language model itself. Understanding how AI interacts with external tools and data sources is critical.
One level was particularly tricky because the bot generated convincing but fake outputs - hallucinated "flags" that looked real but weren't. Recognizing when a bot is hallucinating vs revealing actual secrets is a skill that only comes from practice. Don't trust the first answer that looks like a flag.
Overall, I finished in about 2.5 hours with a perfect score. The exam is one of the most fun and realistic AI security certifications available. It feels like a real engagement, not a test.