Guardrail findings
Requires the Pro plan or higher; see Billing. On a lower plan, the Guardrails page offers Upgrade to Pro. Viewing the page needs Admin, or a role with the View guardrails permission.
Guardrails scan answered traces for PII, secrets and denylisted terms. They never block a response. They give you a record of what was said and, with auto-pause, a way to stop an agent whose findings keep repeating.
To check calls while they happen and block them, including with your own safety model, see Use your own guard model.
The checks
You turn each check on per agent:
| Check | What it flags |
|---|---|
| PII detection | Emails, phone numbers, card numbers and similar identifiers. |
| Secrets & credentials | API keys, bearer tokens and other credential-shaped strings. |
| Denylist | Your own banned terms or phrases, per agent or workspace-wide, as plain text or a regex. |
Each check has two modes:
- Flag: counts toward auto-pause.
- Log: records the finding but doesn't count toward auto-pause.
There is no block mode, because the scan runs after the response is delivered.
Set up guardrails for an agent
Create guardrail opens a wizard. Steps you don't need, such as denylist terms when Denylist is off, are skipped.
Choose the agent
The agent whose traces to scan. You can also add Tags, which are saved onto the agent. See Tags and bulk actions.
Choose checks, and a mode for each
Turn on any of PII, Secrets & credentials and Denylist, and set each to Flag or Log.
Add denylist terms (only if Denylist is on)
Enter a term, mark it as a regex if it is one, and choose this agent only or every agent in the workspace. Each term is saved as soon as you add it, not with the final Save.
Turn on auto-pause (optional)
Set Pause after a number of findings. When the agent's Flag-mode finding count reaches it, the agent stops accepting calls.
Review, then save
Review each check's state and mode, the denylist term count and the auto-pause setting, then save.
To change an agent's setup, choose Edit checks on its row. The wizard opens at the checks step.
The guardrails table
Each agent with a guardrail has one row:
| Column | What it shows |
|---|---|
| Agent | The agent, with In review — paused under its name if auto-pause stopped it. |
| Tags | The agent's tags. |
| Scanning | live while scanning is on, dormant while it is off. |
| Check | Which checks are on. |
| Check status | ready when at least one check is on, dormant when none are. |
| Findings | The number of findings. With auto-pause on, a line such as 2/5 before auto-pause shows how close the agent is to being paused. |
Filter by Check type, Check status or Tag, or search agents and tags by name.
Reading findings
Click a row's finding count to see recent matches: when, which check, input or output, mode, the matched value (masked), and a link to the call's trace.
Turn scanning or auto-pause off
Scanning and the check policy are independent. Turning scanning off keeps the checks configured for later.
From a row's ⋮ menu:
- Disable agent scanning stops scanning this agent's traces.
- Disable auto-pause keeps flagging findings but never pauses the agent.
When off, these read Enable agent scanning and Enable auto-pause.
Auto-pause
Auto-pause is off by default. When an agent's Flag-mode finding count reaches the threshold, its calls are refused until someone resumes it. Log-mode findings never count.
Turning auto-pause back on doesn't count findings from while it was off.
Resume a paused agent
A paused agent shows In review — paused under its name. Open its row menu and choose Resume from review. This resets the count toward the threshold to zero. The total findings and their records are kept.
Change many agents at once
Select rows to change several agents together:
- Scanning: on or off.
- Checks: each check on (flag), on (log only), off, or unchanged.
- Auto-pause: on with a Pause after number, or off. Turning it off keeps each agent's number for later.
- Resume from review: unpause paused agents.
- Edit tags: add, remove, replace or clear tags.
You see a preview of what will change and what will be skipped before anything is applied. See Tags and bulk actions.
Scan limits
- Scans run on an interval, so a finding and any auto-pause usually land within about a minute of the call.
- Scanned text is capped at about 50 KB of prompt plus completion. A finding
from a truncated scan is marked
scanned_truncated. - A timeout applies to the whole scan of one trace.
- If one check fails, it is logged and the other checks still run.