Toronto Java Users Group

In 2026, AI agents made headlines for the wrong reasons. A coding agent deleted a company's production database in nine seconds. An agent being tested on the open web submitted an invented murder tip through a police department's online form. AI agents in a security test escaped their sandbox, coordinated with one another, and broke into another company's systems. Were these agents "rogue"? Not in the way the headlines suggest. None of them was trying to cause harm. They were trying to finish their task, and nothing stopped them.

For this talk, no prior experience with AI agents is needed. We'll start from the basics: what turns an LLM into an agent, and why the harness around the model (tools, instructions, skills, permissions, hooks, human approval, monitoring) matters as much as the model itself. Along the way, we'll touch on MCP, Agent Skills and A2A. Then we'll walk through recent incidents, map each one to the piece of the harness that was missing, and look at the controls that would have stopped it. You'll leave with practical questions to ask before putting agents into production, so you don't lose control of them.

Speaker bio
Mohamed Mansour is a Principal Engineer and Architect with more than 25 years of experience across big tech, enterprise and government, including Amazon, IBM, Loblaw Digital and the Ontario Public Service. He has programmed in Java since 1996, the year Java 1.0 was released. His earlier research won a Best Paper Award at ACM/IEEE MoDELS, and his empirical study of requirements engineering at a large financial institution is still cited today. He taught introductory and advanced software engineering as a faculty member at the University of Toronto, including Java to hundreds of students, and now serves on Northeastern University's Program Advisory Committee for its Master's in Software Engineering Systems. As an architect, he designs and modernizes large-scale systems; his current focus is AI-driven legacy modernization and deploying AI agents safely, including contributions to the open-source DeepWiki-Open project and Daniel Oh's Agentic AI Java Workshop.

He is available for team talks, workshops, and architecture reviews of AI agent and legacy modernization projects.

About the venue
Free Times Cafe has bistro-style seating and a full food and drink menu. Please consider helping to support the venue by planning to have supper during the talk.