The sandbox is supposed to provide a boundary between an AI agent and the systems we care about, but recent incidents demonstrate that it is often less solid than we think. As AI agents gain autonomy, broader tool access, and more powerful credentials, they are finding ways to move outside their environments, where they can reach real systems, exploit production servers, and cause serious damage.
In this talk, join two GitHub forward-deployed engineers as we examine how agentic systems cross boundaries, including through sandbox escapes, credential inheritance, misconfigured networks, ambiguous instructions, unsafe tools, and trust relationships between agents. Working with GitHub Copilot, attendees will learn techniques they can apply in any agentic environment to evaluate and secure sandboxes and implement practical controls for safer agentic systems.
You will learn:
How to understand the misconfigurations that enable AI agents to escape their environments
How to identify how agent capabilities create new attack paths across containers, networks, package registries, tools, and other agents
To design layered controls for agentic systems, including identity isolation, tool allowlists, approval gates, and monitoring