Security primitives
- Tenant isolation: hard separation of data and execution contexts.
- Identity and access: SSO integration, RBAC/ABAC for tools and knowledge.
- Secrets: centralized secret management; never exposed to the model.
- Audit logging: immutable logs of access, actions, and approvals.
- Retention: configurable policies for prompts, traces, and knowledge artifacts.
Prompt-injection and untrusted content
- Treat retrieved documents as data, not instructions.
- Separate system policies from user content and source content.
- Allowlist tool usage; never execute actions based solely on document directives.
- Use structured extraction when reading untrusted sources (facts-only parsing).
Operational controls
- Approval gates for privileged actions and data exports.
- Redaction rules for sensitive fields in logs and responses.
- Dependency allowlists (domains, APIs) for external access.