Defending AI Agents

A course for data scientists

AI agents now send email, run code and change records. Keeping them safe is less about clever prompts than about problems data scientists already know: detecting rare events, choosing a threshold under a false-alarm budget, noticing when production data stops looking like your test set, and judging whether a policy change helps before you turn it on.

This course teaches ML engineers, statisticians and data scientists the security half, by having you build every component yourself on a laptop, and then working through what changes at the scale of a large organisation.

What you will build

• A small tool-calling agent, and the attack that breaks it
• A gateway that intercepts every tool call before it runs
• Prompt-injection detectors, with thresholds set from measured false-positive budgets
• An evaluation harness that tests on data shaped like real tool calls, not prose
• A policy engine that allows, refuses, or holds an action for a person
• A tamper-evident, signed audit log, and replay of a policy change over past traffic

Who it is for

People who are comfortable with Python, probability and machine learning. No security background is assumed.