Monitoring, red teaming & incident response

Lesson 8 of 8 · 30 min · part of the AI Security Professional certification

No defense is perfect, so mature AI security programs assume breach: they red-team before launch, monitor in production, and have an incident-response plan specific to AI failures.

• Red teaming — structured adversarial testing (injection, jailbreaks, data leakage, harmful output) before and after each release.
• Guardrails — input/output classifiers and policy checks as one layer of defense in depth.
• Monitoring — log prompts, tool calls, refusals and spend; alert on anomalies such as spikes in tool use or leaked canary tokens.
• AI bill of materials — track which models, datasets and prompt versions are in production.

Interactive exercise: Respond to an AI incident

Interactive exercise: Course review