An AI system that works is not automatically one you can trust. AI introduces risks that ordinary software doesn't have: prompt injection, poisoned training data, leaked sensitive information, biased or harmful outputs, and models that behave unpredictably. This module covers AI security and threat modeling, along with the ethics and societal risks you need to weigh when you build and ship AI.
Once a system is live, you also need to see what it is doing. Observability means tracing each request, monitoring cost, latency and quality, and debugging failures with real data. Combined with guardrails and ongoing evaluation, it is how you keep an AI product safe and reliable over time.
These courses are long, so treat them as a reference: watch the parts you need, then come back for the rest. You've reached the end of the modules. Continue on to the Projects to put everything into practice.
Why securing an AI system differs from securing a normal app, how to threat-model AI (from poisoned training data to supply chain attacks), and how to defend against prompt injection, adversarial examples and sensitive data leaks.
A broad look at the ethical side of AI: neural networks and datasets, black-box limits, bias, philosophical foundations, AI risk, and the impact on healthcare, education, finance, employment and policy, with guest speakers.