Secure, kernel-enforced sandbox for AI agents, MCP and LLM workloads. Capability-based isolation with secure key management and blocking of destructive actions in a zero-trust environment.