Security and deployment
An agent can read and write files, execute commands, access networks and invoke external systems. Security is a boundary across identity, models, tools, data and deployment.
Principles
- grant the minimum directory, tool, model and network scope;
- separate read/search operations from write/delete/publish actions;
- require human approval for high-risk operations;
- fail closed when an unattended task cannot obtain approval;
- never place API keys, pairing secrets or credential-bearing logs in Git or chat;
- isolate projects, customers and execution environments;
- treat web pages, documents, repositories, artifacts and MCP services as potentially untrusted.
Deployment options
| Pattern | Description |
|---|---|
| Personal local | Credentials and repositories stay on an individual device |
| Team managed | Platform services are centralized while execution remains on member or dedicated nodes |
| Private deployment | Platform, HiPMM, IM and execution nodes run in an organization-controlled environment |
Before procurement, document data flows, key rotation, approval rules, network reachability, retention, deletion, backup, audit and failure recovery.
WARNING
No serious agent system is “zero risk.” Define the threat model, implement controls and validate high-risk workflows in the real deployment environment.