Four Ways to
Poison an AI Agent.
Everyone's connecting agents.
Who's governing them?
The excitement around agentic AI is electric, but not enough people are talking about tool poisoning, memory poisoning, over-broad permissions, malicious MCP servers, and the reality that traditional API security was not designed for autonomous agents.
We'll cover practical techniques organizations should be using today — least-privilege access, tool governance, runtime controls, memory hygiene, auditability, and trust boundaries.
Paddy will also share how Atmosoft is approaching these challenges and why the next major challenge isn't connecting agents — it's governing them safely once they are connected.
A timely conversation as the industry races to make agents more capable, while many organizations are only beginning to think about how to operate them safely in production.
Paddy Gonzalez
Paddy Gonzalez is Managing Partner at Atmosoft, where he is building the future of AI defense and private-AI infrastructure. His background spans security architecture, credit and authentication systems at AVB, scrum and delivery leadership at Capgemini and W3, and a degree in MIS from Montclair State. He works at the intersection of agentic AI, MCP, and enterprise security — helping organizations connect agents to real systems without losing control of them.