Remy Guercio discusses the boundary and permissions of an AI agent sandboxAn AI agent sandbox is an isolated execution environment that limits which files, processes, networks, credentials, and external systems an agent can access., and the risks of placing credentials inside the agent’s environment.
Using a gateway demonstration, Remy Guercio explains how network identities and tags can govern model access, quotas and visibility into requestsAn agent permission boundary defines and enforces which resources and actions an automated agent is allowed to use.. These are the capabilities described in the talk, not a guarantee that every agent action is observable.
The discussion covers access policies, limitations of transparent routing and the added difficulty of interpreting agents that execute generated code.
Watch on YouTube
