AI interactions

Overview

The Shield360 browser extension records every prompt a person submits on a supported AI site, not only the ones a guard flagged. Each record says who sent it, when, to which app and model, from which kind of account, how long it was, which types of file went with it, which kinds of sensitive data it held, and what the guard did.

Records appear under Govern → Records → AI interactions. You can search and filter them, export them as CSV or JSON Lines, send them to your SIEM, and keep them under a legal hold.

Prompt and response text is not recorded unless your organisation chooses to store it. See Prompt and response text.

Prerequisites

• The browser extension installed in managed Chrome or Edge, and connected to the environment.
• To open or export stored text: the Read stored AI interaction text permission. Administrators and compliance officers have it.

Supported sites

The extension reads the composer on these sites. On each one it records interactions and applies prompt protection (warn, mask or block sensitive data).

Site	Host	Model shown	Account type	Response capture
ChatGPT	chatgpt.com	Yes	From the plan (Free, Plus, Pro: personal; Team, Enterprise: company)	Yes
Claude	claude.ai	Yes	From the plan	Yes
Gemini	gemini.google.com	Yes (mode)	Unknown	Yes
Microsoft Copilot	copilot.microsoft.com	Yes (mode)	Personal (consumer accounts only)	Yes
Microsoft 365 Copilot Chat	m365.cloud.microsoft/chat	No	Company (work accounts only)	Yes
Perplexity	perplexity.ai, www.perplexity.ai	Yes	Unknown	Yes
Le Chat (Mistral)	chat.mistral.ai	No	Unknown	Yes
DeepSeek	chat.deepseek.com	No	Unknown	Yes
Grok	grok.com	Yes	Unknown	Yes
Meta AI	meta.ai, www.meta.ai	No	Unknown	No
Poe	poe.com	Yes (bot name)	Unknown	Yes

Other AI sites in the catalog, such as you.com and character.ai, are allowed, warned or blocked by the AI tools policy, but prompts on them are not read or recorded.

Note

On DeepSeek the send control has no label, so only sends by pressing Enter (or a form submit) are recognised. Clicking the send arrow sends the prompt without a record or a check.

Caution

Sites change their pages without notice. When the extension cannot find a site’s composer it reports the site as an adapter miss in its heartbeat, shown under Discover → AI usage → Computers. Until the adapter is updated, prompts on that site are neither checked nor recorded.

What a record holds

Field	Description
When	The time the prompt was submitted
Person	The managed browser profile’s email
Browser	The extension’s device id (browser:<id>)
App and site	For example gemini-web on gemini.google.com
Model	The model or mode the page shows as selected, as its label reads. Empty when the page does not show one
Account type	company, personal or unknown, from what the page shows. A Team or Enterprise plan reads as company, even when it is another organisation’s. AI account identity tells those apart
Prompt length	Characters in the prompt
Attachments	How many files were picked, dropped or pasted since the last prompt, and their types (extensions). File names and contents are never recorded
Kinds of sensitive data	For example email, api_key. Never the values
Decision	Allowed, warned and not sent, warned and sent anyway, sent with data masked, or blocked
Sent	Whether the prompt went to the AI site
Policy version	The enforcement policy the extension decided under
Text stored	Whether prompt or response text is stored for this record

Prompt and response text

The extension sends prompt text, and the response once the page finishes writing it, only when both of these are true:

1. The organisation stores prompts and responses: Settings → Data retention → Prompts and responses is on for the environment. See Data retention.
2. Content capture for the person is Full in the collection policy.

Otherwise the text never leaves the browser, and only the record is sent.

When text is sent:

• The extension replaces every sensitive value it detects with a placeholder ([EMAIL_1]) before the text leaves the browser.
• Shield360 checks again at ingest. It replaces credentials with a type and fingerprint and, while Mask personal data is on, masks personal data.
• A record holds at most 32,000 characters of prompt and 32,000 of response. Longer text is cut, and the record says so.
• A person narrowed below full content capture has their text dropped at ingest, whatever their browser sends.

Turning Prompts and responses off stops new text from being stored. Text already stored stays until it ages out.

Retention

What	Kept for
The record	365 days
Prompt and response text	90 days, then removed from the record
Copies under a legal hold	Until the hold is released

Search and filter

Govern → Records → AI interactions uses the time range in the page header. You can narrow the list by:

• Search: matches the person, app, site or model. It never searches prompt text.
• Decision and account type.
• With stored text only.
• App, person or kind of data: choose one in the breakdown panels or in a row.

The table shows the 200 most recent matches. Exports include every match.

Open stored text

Where text is stored, a row shows View text. Opening it shows the prompt and the response, with sensitive values already replaced.

Each opening is recorded in the audit log as compliance.ai-interactions.content.reveal with the record’s id. The text itself is never written to the audit log.

Export

Export CSV and Export JSON download every record matching the current filters. JSON is in JSON Lines format, one record per line, which most eDiscovery tools import directly.

Exports hold the record fields only. To include stored text, select Include stored text. This needs the Read stored AI interaction text permission, and is limited to 10,000 records per export.

Every export is recorded in the audit log as compliance.ai-interactions.export, with the time range, the format and whether text was included.

API

curl -X POST "https://<your-shield360>/api/compliance/ai-interactions/export" \
  -H "Content-Type: application/json" \
  -H "Cookie: <session>" \
  -d '{
    "timeLimit": { "start": "2026-10-01T00:00:00Z", "end": "2026-10-08T00:00:00Z" },
    "decision": "blocked",
    "format": "json"
  }'

Route	Permission	Purpose
POST /api/compliance/ai-interactions	View data protection events	Summary, breakdowns and the 200 most recent records
POST /api/compliance/ai-interactions/export	Export data protection evidence, plus Read stored AI interaction text with includeContent	CSV or JSON Lines
POST /api/compliance/ai-interactions/content	Read stored AI interaction text	One record’s stored text, by eventId

All three read the current project’s environment. The routes belong to the Compliance module.

SIEM

Select AI interactions (ai_interaction) as a category on a destination under Settings → Integrations → SIEM. Each record is sent as an OCSF API Activity event. Prompt and response text is never sent to a SIEM. See SIEM export.

Legal holds

AI interactions are a legal-hold source. A hold on a person copies their records, including stored text, into the hold. Held copies are not removed when the record or its text ages out.

Personal data requests

An access or erasure request for a person includes their AI interactions, found by the profile email.

Related

• Data retention
• Enforcement policy
• AI account identity
• SIEM export
• Violations