AI Safety, Security and Governance

A model that works in a demo can still state false facts with confidence, follow instructions hidden in a web page, leak personal data into logs or treat some groups of users worse than others. None of these failures is solved by a better prompt alone. They are handled the way engineers handle other risks: by understanding the mechanism, measuring how often it happens and putting several layers of defence around the model.

You will start with hallucinations: why models confabulate and how grounding, citations and abstention reduce it. Then you will see why prompt injection is so hard to stop once an agent mixes private data, untrusted content and a way to send data out, and how guardrails, privacy controls and bias evaluations work in production. The module ends with the bigger picture: what alignment and interpretability research is trying to do, and how risk-based regulation such as the EU AI Act, model cards and audits shape what you are allowed to ship.