Governance

Organization-wide policy that detects personally identifiable information in both agent inputs and outputs. Inbound data is scrubbed through the centralized sanitization pipeline before model invocation. Outbound data is validated at the output layer before transmission to external APIs. Covers SSN, credit card numbers, email addresses, phone numbers, physical addresses, and other PII patterns defined by compliance.

Automatically approves JIT access requests when the requested permission is data:read and the target resource sensitivity is public or internal. Grants are time-limited to 60 minutes.

Requires human approval for any JIT request involving data:write or data:delete permissions. Escalates to the security admin for review before granting access.

Controls access to memory instances with confidential or restricted sensitivity. Denies access by default unless the requesting principal has a security-admin or org-admin role. Agent-operators may request access through an approval workflow with a 30-minute TTL.

Organization-wide execution guardrails that enforce cost limits, error rate thresholds, and recursion depth caps on all agent runs. Designed to prevent runaway agents and contain blast radius.

Escalation policy that automatically kills agent sessions after 3 or more consecutive errors or when anomaly detection scores exceed 0.9, indicating abnormal agent behavior.

Controls how memory instances can be shared across teams. Public memory can be shared freely, while internal-sensitivity memory requires approval from the memory owner or a security admin.