Enforcement actions
Guardrails decide what counts as sensitive. Enforcement actions decide what happens when something is caught. You set them on the Guardrails page, on the Enforcement Policy tab and in the Enforcement column of Guardrail Catalog.
The three actions
Every guardrail has one of these actions:
- Nudge: the Security Check warning appears, with Sanitize Prompt, Edit Prompt, and Send Anyway. The person can go ahead, and the override is recorded.
- Block: the Blocked by Policy warning appears, without Send Anyway. The prompt can’t be sent until the sensitive data is removed. Sanitize Prompt and Edit Prompt stay available, so nobody is stuck.
- Monitor: no warning at all. The prompt is sent and the event is recorded silently. Use it when you want visibility before you decide to act.
To see what people see, see When a prompt is flagged.
Defaults by severity
You don’t have to set an action on each guardrail. Each guardrail has a severity, and the severity picks the default action:
| Severity | Default action |
|---|---|
| Critical | Block |
| High | Block |
| Medium-High | Nudge |
| Medium | Nudge |
| Low | Nudge |
The Enforcement Policy tab shows this table, with how many active guardrails have each severity.
Change the action for one guardrail
- On the Guardrails page, open Guardrail Catalog.
- In the Enforcement column of the guardrail, choose Nudge, Block, or Monitor. To follow the severity default again, choose the Default option.
A guardrail you changed shows Overridden under its action. It also appears in the Overrides card on the Enforcement Policy tab, where Reset puts it back on its default. That way, exceptions to your own policy stay visible.
Before you turn on a block
When you set a guardrail to Block, a dialog titled Enable hard block? shows what the block would have caught recently, for example In the last 30 days, this would have blocked 142 prompts from 37 users. You can then:
- Click Enable Block to block right away.
- Click Start as a nudge for 14 days to warn first and block later. See Soft launch a block.
- Click Cancel to keep the current action.
Treat a high number as a warning sign. Either the data really is pasted into AI tools a lot, and the block does its job, or the guardrail fires on things that aren’t sensitive, and you should fix that first. If you’re unsure, start with a soft launch.
Stricter rules on riskier AI tools
Not every AI tool carries the same risk. Risk-aware enforcement, on the Enforcement Policy tab, lets you pick a risk level. When a prompt is headed for an AI tool at or above that level, nudges become blocks. The same data can then nudge on your approved enterprise tool and be blocked on a risky free one, without two sets of guardrails.
Good to know:
- Only nudges escalate. A guardrail set to Monitor stays silent, and one set to Block already blocks.
- Each AI tool’s risk level comes from its report card on the Guarded Sites page.
Check whether it’s working
On the Events page, the Overrides card shows the share of flags people sent anyway, and which guardrails they override most.
That list is your tuning queue. A guardrail near the top is one your team routinely works around. Either it fires on data that isn’t sensitive, and the guardrail needs fixing, or the data genuinely needs to be sent, and an exception is a better fit than a daily obstacle.