Guardrails and privacy
Guardrails redact sensitive values from what CortexHub stores. Redaction is storage-only: it shapes the activity record that is kept for spans and audit, never the live call. The agent always receives the full call and the full response.
Data handling
Open Security -> Data handling in Org space:
- Store tool-call payloads: whether tool-call arguments and results are kept in CortexHub at all. Turn it off to keep payloads out of storage entirely.
- Privacy mode:
redact(run the redaction pipeline over stored payloads) oroff(store raw). - Fail mode:
closed(drop the payload if redaction errors) oropen(store it raw).
This applies to every member in the org.
Built-in detectors
The Data handling tab lists the built-in secret detectors available in the CortexHub catalogue. CortexHub manages the catalogue; your org turns each detector on or off per surface with the switches shown next to it. Each detector applies on the surfaces where it is enabled:
| Surface | What it covers |
|---|---|
| Files | Secrets in uploaded document chunks at ingest |
| Tools | Secrets in tool-call arguments and results (respects privacy mode) |
| LLM | Secrets in co-worker chat, Brain, and prompts sent to the model |
Secrets on the LLM surface are scrubbed from the prompts sent to the model, so they are not persisted in co-worker text.
Custom guardrail patterns
Add your own regex patterns under Security -> Runtime guardrails -> Custom guardrail patterns. Org patterns apply in addition to the built-in detectors.
Group guardrails
A group may add stricter custom patterns only; it cannot relax the org floor. Configure them on a group's Guardrails tab under Groups.
What members see
Members do not edit guardrail patterns or the org data-handling settings. They experience redaction as masked fields in their session views and exported telemetry. (Members can, separately, tighten the access policies on their own toolkits and personal files; see Policies and approvals.)
Data handling and the policy floor
The Data handling tab covers redaction and privacy. CortexHub also has a separate policy floor that governs approvals and access (require approval, step up, redaction, or deny). Like guardrails, the floor is tighten-only: it can only make access stricter, and group or tool policies cannot relax it.
You manage the floor through plain-English rules and compliance profiles on the Policies page. Curated profiles (such as HIPAA, PCI DSS, and Financial operations) apply as fixed bundles, and you can create and edit your own custom profiles, adding or removing rules to fit your org. See Policies and approvals for how profiles apply and merge.