
AI Security Architecture: Production-Grade Defense
Udemy ยท English
About this course
- Language
- English
- Category
- Provider
- Udemy
- Data updated
Your language models have left the chat window. They read your documents, search your knowledge base, query your databases, and โ through the agents your teams are shipping now โ open tickets, move money, and change records. Every connection is capability. Every connection is also attack surface of a kind the classic security playbook was never written for: a component that can be persuaded, whose context mixes your instructions with text written by strangers. This course teaches AI security as the systems discipline it is. You will start with why AI systems fail differently โ the probabilistic attack surface, the blast radius that grows the moment a chatbot can act, and real incident case files dissected down to the control that would have stopped each one. Then you will threat-model LLM and agent systems properly: assets that include the model's attention itself, adversaries that include the authors of the content your system reads, trust boundaries drawn honestly, and the OWASP LLM Top 10 used as a completeness checklist behind a scored risk register. Then the attacks, one family at a time, with their defenses built โ not described. Direct and indirect prompt injection and the jailbreak families, countered by a layered defense stack you will run against a live payload suite. Data leakage through system prompts, shared context, and logs, countered by PII detection and redaction pipelines placed before data reaches a model. The supply chain โ poisoned models, backdoors, untrusted datasets, adapters, and plugins โ countered by provenance checks, signing, pinning, and admission control. Agents scoped to least privilege per task, tool output treated as untrusted input, sandboxing and approval gates that hold under fatigue, and delegation that narrows authority at every hop. Retrieval secured from inside: corpus-poisoning defenses, index integrity, and per-user access control enforced in the search itself.











