Most AI incidents aren't bugs. They're design flaws. Picture a support assistant that refunds $5,000 on a $50 order because the system let the model pick the amount. This course teaches you to find flaws like that while you design the system, before they turn into incidents.
One test runs through every lesson: assume the model is wrong. What keeps the system safe? You'll map an AI system and find where trust changes, sort the controls that enforce security from the ones that only make mistakes less likely, and trace whose authority each action carries. Then you'll threat-model a complete system and check that your fixes hold.
Built for developers working with LLMs, RAG and agents, and for the AppSec engineers and architects who review them. LLM Security Fundamentals and MCP Security Fundamentals help, but aren't required.
Curriculum
Five sessions and a short wrap-up: 18 video lessons, about 2 hr 50 min, with a quiz after each session and a final assessment. Watch in Arian's voice or with an AI narrator, with subtitles in English, Spanish and Portuguese.
The Mindset
Why the model can't be your security: to a model, data and instructions are one stream of text. The core rule (the model proposes, your code decides), the test of assuming the model is wrong, and the map of where controls belong. Three lessons.
Understanding the System
How an LLM app is built, from the model and its context to retrieval, memory, tools and MCP. Follow one request through LeakyCart, a store with an AI support assistant, and mark where trust changes. Three lessons.
Securing What Goes In and Out
Tell locks (checks in your code that hold when the model is wrong) from extras (anything that makes a wrong model less likely). Control what the model is allowed to see, and stop its replies from leaking data. Three lessons.
Securing Identity and Actions
Authorize against the person, not the agent. Decide who sets each tool value, when an action needs approval, and how to limit the damage when something goes wrong. Then trace whose authority a request carries, hop by hop. Four lessons.
Threat Modeling
A six-step method: draw it, trace it, ask the map's questions, place the locks, prove one, write what's left. A worked example on a marketplace seller assistant, then how to gather evidence and re-test your own fixes. Three lessons.
The Whole Course in One Page
Six ideas, six artifacts, and what to do in your first week with an LLM feature you own. One lesson.
Earn a certificate
Finish the course and earn a verifiable certificate — issued by AI Security University with its own public verification page, and shareable straight to your LinkedIn profile.
- Public verification page
- Counts toward 4 CPE credits
- Open Badge compatible
- Add to LinkedIn
To earn it: Complete all course lessons and pass the final assessment with 80% or higher.
See a sample credentialYour instructor
Arian Gharedaghi
Arian is an AppSec & DevSecOps Engineer in the fintech industry. He runs the YouTube channel Sec with Arian, where he teaches AppSec in plain language, with a focus on deep understanding and helping people become dev-native security engineers.
He contributed to the Secure-by-Design section of the MLSecOps Practical Reference Guide.
Find the design flaws before they become incidents.
Free and self-paced. Join the community and we'll let you know the moment it opens.
This Week in AI Security
The AI security stories that matter, every Sunday. Free on Substack.