research
docs://contents>
docs/security-and-isolationdoc 20

Security and Isolation

Security in this plane is mostly a property of construction: tools only execute inside the MCP sandbox, the sandbox is ephemeral and restricted, and the entire plane runs inside your perimeter. This document specifies the sandbox lifecycle, the threats the model answers, and the sovereignty posture.

The sandbox lifecycle

provision ──▶ fetch tool ──▶ execute ──▶ destroy
   │            │             │
restricted    verified      network-limited
container     tool code     filesystem-restricted
              download
Stage Semantics
Provision An isolated container is created for this execution
Fetch The MCP tool's code is downloaded into the container
Execute The tool runs with restricted network and filesystem access
Destroy The container is destroyed immediately after execution — nothing persists

The unit of isolation is the call: one container per tool execution, no reuse across steps, no state carried between executions.

Threat matrix

Threat Stopped by
Prompt injection attempting to exfiltrate data Network-restricted sandbox — the exfiltration call cannot leave
A compromised tool writing to the host Filesystem restriction — the write has nowhere to land
An agent inventing a tool or parameters The input schema gate — the call is never made
A disallowed action on an external system Cedar evaluation — DENY halts before execution
Credential capture through logs Credentials never enter the Evidence Ledger; runtime injection only
Data leaving the perimeter Air-gapped deployment — see below

Data sovereignty

Environment Posture
Staging The plane runs locally in your staging environment
Production Deployed into your VPC through containerized templates
Egress Zero — your workflows, tools, and data do not leave your perimeter

Your SOPs, Cedar policies, MCP tool definitions, and evidence all live and run inside the perimeter. The delivery model is white-glove: the plane comes to you; nothing of yours goes to it.

Credential handling

Credentials — LLM API keys, tool authentication — are yours and stay yours. They are injected at runtime into the execution context, and the redaction rule is structural: the ledger records that a call was made with evaluated parameters; it never records the credential material used to make it. An auditor can verify every decision the plane made without ever handling a secret.