Guard

Guard is a set of safety checks the platform runs on your inference requests. It looks at what goes into a model and what comes back, and acts when something looks unsafe, such as an attempt to hijack the model's instructions. The first check is prompt-injection detection, which scans requests for common ways people try to override a model's behaviour.

You will find Guard under Services > Guard.

Turning Guard on

Use the Enabled switch in the top right of the page. It takes effect the moment you flip it, so there is no separate save step for turning the feature on or off. While it is off, requests pass through without any checks.

What happens when a check fires

When a check detects something, what the platform does next is up to you:

• Allow (log only). Let the request through, but record that the check fired. Useful while you are getting a feel for how often it triggers.
• Flag. Mark the request as flagged and let it continue.
• Redact. Remove the offending part of the content before the request continues.
• Block. Stop the request and return an error. Nothing is sent to a model.

You set this with the Action control. The Threshold sets how confident a detection has to be before the action is taken, from 0 to 1. Change either one and press Save to apply it.

Detection groups

Prompt-injection detection is organised into groups, each covering one technique, such as direct instruction overrides, prompt-extraction attempts, or jailbreak phrasing. A few groups catch evasion tricks, like content that has been base64-encoded, spaced out letter by letter, or scrambled, and these have no fixed patterns of their own.

The table lists every group. For each one you can:

• Turn it on or off with the switch. This saves straight away, so you can tune exactly which techniques are scanned without pressing Save.
• See its patterns. Click the pattern count to open a dialog listing the exact expressions that group matches against, with a short description of each.

Turning a group off means the platform stops scanning for that technique. The others keep working.

Metering and records

Each check is metered, so it shows up in your usage. Every time Guard acts, it writes an entry to the audit log under Platform > Audit logs. As everywhere on the platform, those entries record what Guard decided, never the content of your request.

Good to know

Guard must be enabled. If the Enabled switch is off, no checks run, regardless of how the groups are configured.

Changes apply to your whole organization. The action, threshold and group settings are set once per organization and apply to every request you send.

Block is final. A blocked request never reaches a model, so it is not billed for inference.