GenAI safety is a layered defense at three boundaries: input (what reaches the model), model behavior (what it's trained/instructed to do), and output (what reaches the user or downstream systems). No single layer is sufficient.

The threat taxonomy

• 1. Prompt injection
• 2. Jailbreaks

Key distinction: prompt injection = hijacking the system's instructions. Jailbreak = hijacking the model's trained behavior. Both can co-occur. The defenses differ.

A clean canonical sentence to have ready: "I think about prompt injection as instruction hijacking — direct via user input, indirect via retrieved content — and jailbreaks as alignment circumvention. They're related but the mitigations are different."

• Layer 1 — Input filtering
• Layer 2 — Model behavior (alignment + system prompts)
• Layer 3 — Output filtering
• Vertex AI safety filters — the specifics
• The interview synthesis
• Common probes you should be ready for