RESEARCH ARCHIVE

Notes from the
adversarial edge.

Editorial field notes on software, AI, and security. These educational articles are separate from validated vulnerability advisories.

R—01Agentic AI

The confused deputy is back. This time, it has tools.

Why agentic systems collapse trust boundaries—and a practical framework for constraining tool use under hostile context.

R—02Model evals

Moving from static benchmark scores to security signals you can actually operate.

R—03Prompt injection

A threat model for retrieval systems that assumes every document is adversarial.

R—04Agent security

A faster way to find the dangerous edges of an AI agent before writing attack scenarios.

R—05Agent security

Why confirmation dialogs fail—and how to bind human intent to the action an agent actually takes.

R—06Red teaming

A compact structure for turning surprising model behavior into an engineering decision.