02 / Writing
The Ultimate Guide on Prompt Injection
Published on Algolia's blog in July 2024: a practical guide to defending LLM apps, with code you can actually paste. It starts with the question most teams skip: do you even need an LLM?
Context
Prompt injection is easy to describe as a spooky AI problem. The useful version is more specific: what can the model influence, what belongs to the application, and where should the boundary hold?
The problem
The guide had to explain the risk without turning every model output into a catastrophe—or every mitigation into a magic trick.
Work involved
I researched and authored the guide, turning security concepts into a structure technical readers could use.
Approach
- Separated model behavior, application boundaries, testing, and controls so they did not blur together.
- Used practical examples and adversarial thinking to show how inputs can steer a system.
- Connected mitigations to layered defense and least privilege.
What this work shows
The subject has to stay accurate while remaining usable. That is the work here: research, information architecture, technical security communication, long-form writing, and editing.
Technical or communication skills involved
Outcome
A public guide that gives readers a practical way to think about prompt injection, input separation, adversarial testing, layered defenses, and least privilege.