New course: AI Security Engineer
Shipping an LLM feature adds an attack surface that most secure coding checklists were never written for. A model can be talked into ignoring its instructions, made to leak the data it was given, or handed enough agency to take actions no one signed off on. AI Security Engineer, now on SecDim Learn, is about writing LLM-powered code that holds up against all of it.
The course is built around the OWASP LLM Top 10, and it treats each item as something you can see in code rather than a category to memorise. It moves through prompt injection, sensitive information disclosure, supply chain risk from third-party models and datasets, data and model poisoning, improper output handling, excessive agency, system prompt leakage, vector and embedding weaknesses, misinformation, and unbounded consumption of model resources.
Each risk comes with the defensive pattern that answers it, and most come with a hands-on challenge on the SecDim platform where you fix the flaw in a running application. You stop a bot that leaks secrets, find and fix a model that executes a command the moment it loads, rein in an assistant that can do far more than it should, and shut down a chatbot that is burning compute on every request.
It is written for developers, engineers, and security practitioners who are already building with LLMs and want to recognise the real attack scenarios before an attacker does. No prior AI-security background is assumed.
Topics covered include:
- Prompt injection, and how to constrain what the model will act on
- Sensitive information disclosure and system prompt leakage
- Supply chain attacks through malicious models and datasets
- Data and model poisoning
- Improper output handling when model output reaches other systems
- Excessive agency, and least privilege for tools
- Vector and embedding weaknesses in RAG
- Unbounded consumption and model denial of service
- An OWASP LLM Top 10 certification challenge to finish
If your roadmap has an LLM on it, this is the secure coding that should go with it.
Check it out on SecDim Learn: Learn - SecDim
