AI Prompt Governance

AI Prompt Governance gives you visibility into what employees send to AI assistants, and lets you act on it before it leaves the browser. The Keep Aware extension inspects a prompt as it is sent — not by reading the page — so a prompt carrying sensitive data can be recorded, paused for acknowledgement, or blocked outright before it ever reaches the assistant.

Prompts are treated as a first-class event type. They appear in your event tables, can be scoped in policies, and support evidence collection like any other activity.

info

Prompt governance is off by default. It must be enabled per tenant before any prompt activity is recorded or governed. See Enable prompt governance below.

Requirements

• Access to your Keep Aware console with permission to change settings and author policies
• Browsers running the Keep Aware extension 3.6.0 or later

Supported assistants

Assistant	Domain
ChatGPT	chatgpt.com
Claude	claude.ai
Google Gemini	gemini.google.com
Microsoft Copilot	copilot.microsoft.com, copilot.cloud.microsoft
Grok	grok.com
DeepSeek	chat.deepseek.com
Perplexity	www.perplexity.ai

Coverage for each assistant is maintained by Keep Aware and delivered to the extension automatically. There is nothing to configure per assistant, and you cannot currently add your own — including internally hosted assistants.

Assistants outside this list are unaffected and continue to be governed by your existing input and upload controls.

Enable prompt governance

Prompt governance has two parts: switching the prompt event type on, then authoring a policy that decides what happens.

1. Turn on prompt logging

1. In the console, go to Settings → Capabilities.
2. Find the Event logging group.
3. Enable Prompt.

Optionally, narrow logging to prompts a policy acted on, by outcome — label, safeguard, or threat. Leaving this unset records every prompt on a supported assistant.

See Event Logging for how this group works in general.

2. Author a prompt policy

Create or edit a policy and scope it to prompt activity. Your existing detection patterns apply unchanged — a prompt is matched the same way as any other content.

Choose the action you want:

Action	What the user sees
Log only	Nothing. The prompt sends normally and the event is recorded.
Pause	The send is held. The user sees your safeguard and can acknowledge to continue, or cancel.
Restrict	The send is blocked. The user sees your safeguard explaining why. There is no continue option.

A paused or blocked prompt always shows the safeguard on screen, whatever style it uses, so the user is never stopped without explanation.

Bypass cooldown

When a user acknowledges a paused prompt, you can honor that acknowledgement for a period of time rather than prompting on every send. Set a cooldown on the policy's safeguard.

Cooldowns apply only to pause. A restrict is always enforced — there is no acknowledgement to carry forward.

Evidence

If your policy collects evidence, the prompt text is captured and stored through your configured evidence integration. Where the integration is configured for encryption, evidence is encrypted in the browser before upload, so plaintext never leaves the device.

Prompt evidence appears in the event side panel and in the conversation timeline on an application's resource Details tab, where prompts and uploads expand into cards showing their captured content and data categories.

note

Raw prompt text is never displayed in the console outside of collected evidence.

Verify it is working

1. Confirm the browser is on extension 3.6.0 or later.
2. With Prompt enabled and a log-only policy in place, send a benign prompt to a supported assistant and confirm a prompt event appears in your events table.
3. Add a pattern that matches test content, set the action to Pause, and send a prompt containing it. The send should hold and your safeguard should appear.
4. Change the action to Restrict and repeat. The send should be blocked and the safeguard shown.

If prompts are not appearing, check in order: the extension version, that Prompt is enabled under Event logging, that your policy is scoped to prompt activity, and that the assistant is in the supported list above.