Synack Guardrails Ebook is an educational resource that provides guidance for organizations evaluating agentic AI-powered penetration testing solutions. The e-book addresses the security challenges posed by AI-powered attackers and offers a framework for assessing AI pentesting tools. The guide covers the translation of traditional human-led penetration testing controls into guardrails suitable for AI agents. It includes a vendor evaluation checklist designed to help organizations assess potential AI pentesting vendors during initial discussions, focusing on governance, policies, and safety approaches. The resource details technical safeguards including command blocklists for destructive actions such as SQL DROP and rm -rf commands, egress filters for scope management, and human-in-the-loop requirements. It identifies four core risks associated with unconstrained AI agents: Uncalibrated Action, Scope Creep, Misinterpreted Context, and Data Mishandling. The e-book is based on Synack's experience in the penetration testing industry and aims to help organizations determine whether their AI pentesting solutions function as trusted partners rather than security liabilities.