AI Safety & Security

Shipping an AI feature introduces risks ordinary software doesn’t have: a new attack surface, sensitive data flowing to new places, and a component that can be confidently, harmfully wrong. This section treats those as core engineering concerns — not compliance paperwork bolted on at the end.

Who owns this

You do. “The model did it” is not a defense — to your users, your security team, or a regulator. The model is a component; the responsibility for the system’s behavior is the engineer’s. This section is the practical version of that responsibility.

In this section

LLM Security Prompt injection, jailbreaks, the LLM attack surface, and defense in depth.

Data & Privacy Where your data goes, handling PII, regulated data, compliance, and logging risks.

Responsible AI Bias and fairness, transparency, human oversight, and knowing when not to use AI.

What you’ll be able to do

Identify the ways an LLM application can be attacked and layer defenses against them; reason about where sensitive data travels and how to protect it; and make honest calls about bias, oversight, and when an AI feature shouldn’t ship.

Prerequisites

AI System Design and AI Agents — the guardrail and tool-security ideas introduced there are expanded here.