An agent that follows malicious instructions may try to send data to an external server, reach other internal systems, or tamper with files. Sandboxing contains the agent in an isolated environment that grants access only to what it needs, limiting the consequences if another control fails. section 4.8.7
Set boundaries around each agent workload:
- Restrict network egress to designated endpoints rather than the open internet. section 4.8.7
- Limit filesystem access to a temporary workspace and clear it after each session. section 4.8.7
- Prevent the agent from starting additional programs and keep credentials outside the sandbox boundary. section 4.8.7 section 4.8.8
Claims agent in an isolated container
ClaimAssist runs agent workloads in isolated containers. An agent can reach the claims database through designated API endpoints, but not the open internet. Its filesystem access is limited to a temporary working directory, and it cannot start other programs. section 4.8.7
Boundaries require judgment: controls that are too restrictive block legitimate work, while permissive controls offer little protection. Sandbox escapes remain possible, and creating a fresh isolated environment for each task can add significant latency. Treat sandboxing as a containment layer, not an impenetrable barrier. section 4.8.7
Use sandboxing alongside other controls. Log agent reasoning traces to support incident investigation, and provide a kill switch that can halt tasks, revoke credentials, and isolate the agent from connected systems. section 4.8.7 section 4.8.8









