CrackML by @ml.with.umang
GenAI interview guides / Prompt Injection & Agent Security Interview Guide
GenAI / LLM

Prompt Injection & Agent Security Interview Guide

Reason about prompt injection as an authorization and data-flow problem, not just a prompt-writing problem.

Threat model

Untrusted instructions can arrive through users, webpages, documents, email, tools, or other agents.

Privilege separation

Keep secrets and powerful credentials outside model-readable context where possible. Gate high-impact actions deterministically.

Data flow

Label trust boundaries, constrain arguments, validate destinations, and prevent retrieved content from silently changing policy.

Testing

Maintain adversarial suites for indirect injection, exfiltration, tool abuse, encoding tricks, and confirmation bypasses.