Agent Influence: Design for Containment

Most security models assume an attacker has to get in first. Autonomous agents don't, they're authenticated, credentialed, and already running inside your environment, with legitimate access to the systems you spent years hardening. The question stops being whether someone breaches the perimeter, and becomes what happens when a trusted agent is influenced into doing the wrong thing.

This session works through what that shift means for teams running n8n across an organization. We'll cover four control points: containment through sandboxed, isolated code execution; credential scoping at the project level rather than the instance level; execution observability detailed enough to work back to why an agent acted, not just what it called; and prompt injection as the primary vector for outside influence.

This is an industry problem, not an n8n problem. We'll use n8n as an honest case study, what the platform gives you today, where the open questions are, and what a durable answer looks like as agent deployments grow.

• Understand why traditional perimeter-based security models fail against autonomous agents that are already authenticated and inside your environment
• Learn the four control points for securing agents in production: containment, credential scoping, execution observability, and defending against prompt injection
• See how sandboxed, isolated code execution can contain what an agent is able to do if compromised
• Discover why scoping credentials at the project level, rather than the instance level, limits the blast radius of an influenced agent
• Identify where n8n's current capabilities stand today, and where the open questions remain as agent deployments grow