Coming to Kindle · 2026
Adversarial Agents
A field manual for AI agent security: securing MCP and thinking like an AI attacker.
Your AI agent can be compromised without a cracked password or a line of malware. One hidden sentence in a document it reads is enough, because to a language model, data and instructions are the same stream. The whole book runs on one idea: an attacker only has to win one of three. You have to hold all three.
Context · what your agent can seeCapability · what it can doChannel · what it can send
Written for: engineers and red-teamers wiring up agents and MCP servers, and the leaders and CISOs who fund and answer for them.
