Docker product manager Rowan Christmas describes an experiment in which a coding agent found sensitive local information, illustrating the limits of relying solely on prompts or harness warnings.
Christmas demonstrates a microVM sandbox intended to isolate the filesystemAn AI agent sandbox is an isolated execution environment that limits which files, processes, networks, credentials, and external systems an agent can access. and control network access and credentialsAn agent permission boundary defines and enforces which resources and actions an automated agent is allowed to use.. The talk also discusses audit trails, agent identity and possible future policy controls.
The demonstration represents the vendor’s account of its own product, and its safeguards should be assessed in the context of each deployment.
Watch on YouTube




