Data & Privacy
Discover AI (Shadow AI) is built to collect usage metadata: which AI tools are used, by whom, how often, and with which files. The browser extension and desktop agents do not send the text of prompts or responses. This page lists exactly what each client sends so you can assess it.
Metadata by design, with three things to know. The browser extension
sends full page URLs and titles of AI tool pages. The desktop agent sends the
command lines of AI command-line tools with every argument value replaced by
<redacted>, so a prompt typed as an argument is not stored. Desktop agents
older than macOS 0.1.30 and Windows 0.1.21 send full command lines, which
ANTS Platform redacts when it stores them. Coding-agent telemetry can include
prompt and response content, but only if you turn on content capture for
that tool.
What each client sends
Browser extension (Chrome, Edge, Firefox)
On pages of AI tools in the Tool Registry and your internal AI tools:
- The full page URL, including any query string, and the page title.
- The AI tool, and the chat or session ID taken from the URL path or query string.
- For a few tools whose URL carries no conversation ID, a conversation marker: a short keyed hash (HMAC) of the first and the latest user message of the chat, used to group requests into one conversation. The key is a random value created for each install that never leaves the device. The message text is not sent. This needs browser extension 0.1.29 or later.
- Prompt submission events: that a prompt was sent, when, the request path it was sent to, the HTTP method and status, and the duration. Not the prompt text.
- The model name, when the request exposes it.
- File uploads: file name, size and type. Not the file contents.
- Time on site, measured while the tab is focused.
On Non-AI Tools, when enabled, the platform keeps only the site address and path, and drops the query string and page title.
The extension also records:
- In Warn mode, each warning banner shown and each dismissal.
- Each acceptance of the employee monitoring notice.
The extension does not read page content, form input, cookies or stored credentials. It records nothing on sites outside the registry.
Desktop agent (macOS and Windows)
-
Which AI apps and command-line tools run, matched against the registry, and how long AI apps are in the foreground.
-
For command-line AI tools: the executable path, the command-line arguments in redacted form, the working folder, the launcher and the names of child processes. Only the executable name and flag names, such as
--model, are kept. Every other argument and every flag value, such as a prompt or an API key, is stored as<redacted>. The agents redact before upload from macOS agent 0.1.30 and Windows agent 0.1.21. ANTS Platform applies the same redaction to command lines from older agents before storing them, and the device panel marks those agents with "Stores full command lines; update the agent". -
Counts from local session files of supported tools: prompt counts, tool-call counts, token counts and the model name. The agent reads these files on the device and sends only the counts and the model.
-
Prompt counts for desktop chat apps: from the composer's send button on Windows, or from the length of the input text on macOS. The text is not read.
-
Files used in AI apps: file name, path, size and type, and how the file reached the tool (file picker, drag and drop, paste, command-line read or write, MCP read or write). Not the file contents.
-
macOS only:
- Clipboard: while an AI app is in the foreground, the agent checks whether a file or image was pasted, and records the file path or the image size. It never reads clipboard text, image pixels or file contents.
- AI network traffic: connection counts and bytes sent and received per
device and host, only for network hosts listed in the registry for a tool
and for AI provider API hosts such as
api.openai.comandapi.anthropic.com. Provider API traffic is recorded with no tool attached, from macOS agent 0.1.32. No request or response bodies. - MCP servers: server names, call counts and connection status. See MCP Server Inventory.
-
Windows, Notion AI only: when a prompt is sent, the Notion window title, which is usually the current page name.
-
Local AI apps and models (macOS agent 0.1.33 and Windows agent 0.1.24 or later), sent with each check-in, only for local AI apps that the Tool Registry gives a local address for, such as Ollama and LM Studio:
- Each app's name and version, and whether it answered.
- Each model's name, family, parameter size, quantization, size on disk, and whether it is loaded in memory.
- For Ollama, the number of requests per day for the last 14 days, counted from the app's request log. The agent reads only the date, method, path and status of each log line, and only the counts leave the device.
The agent asks the app on the device's own loopback address (
127.0.0.1,localhostor::1) only. It never reads prompts or answers of local models. -
Each acceptance of the employee monitoring notice.
Device and user identity (all clients)
Each client checks in about every hour with:
- Device: hardware ID, hostname, serial number (macOS), OS and version, architecture, MDM enrollment, client versions, permission status and health.
- Hardware (desktop agents, from macOS agent 0.1.33 and Windows agent 0.1.24): chip name, installed memory in GB, and whether the device has a neural processor. The macOS agent treats every Apple Silicon Mac as having one. The Windows agent checks the names of the devices Windows lists and sends only a yes when it finds one; the list itself is not sent.
- User: username, full name and email, and where available job title, department and whether the user is a local admin. The enrollment script writes these into managed settings on the device.
ANTS Platform also records the IP address the client connects from.
Coding-agent telemetry (Claude Code, GitHub Copilot CLI, Codex CLI)
These tools send OpenTelemetry data to ANTS Platform: usage events, token counts, cost, model and tool activity. The enrollment script configures this when it finds the tools.
Content capture is off by default. Turn it on per tool under Settings > Deployment & Keys > Coding-agent telemetry > Capture content. With it on, the tool also sends:
| Tool | Content sent with capture on |
|---|---|
| Claude Code | Prompts, tool input and output, and reply text. |
| GitHub Copilot CLI | Messages, and tool call arguments and results. |
| OpenAI Codex CLI | Prompts only. Codex sends no reply text. |
OpenAI Codex CLI always sends tool call arguments and results, even with content capture off. This is how Codex's telemetry works.
The setting reaches devices through the enrollment script and is refreshed on each check-in. See Coding-Agent Telemetry.
What is not collected
Outside the cases listed above, the following are never read or sent:
- The text of prompts or responses in browser or desktop AI tools.
- Prompts and answers of local models.
- Request and response bodies.
- File contents, clipboard text and image pixels.
- Page content, form input, cookies or credentials from websites.
- Keystrokes.
How it is enforced
- Allowlisted fields. Each client sends a fixed set of structural fields. There is no field for prompt or response text.
- Server-side schemas. The ingest endpoint validates every batch against strict schemas and drops any field it does not expect, so an unexpected field never reaches storage.
- On-device checks. The macOS agent checks the output of each local session file reader and discards results that contain oversized strings, which could indicate captured content.
- Registry rules. Detection rules match hosts, paths, bundle IDs and executable names. They cannot point a client at page content. Local AI addresses in the registry are honored only on the device's own loopback address.
- Encrypted queue on the device. Events waiting to be uploaded are stored encrypted on the device, from macOS agent 0.1.33 (with a key in the Keychain) and Windows agent 0.1.24 (with a key protected by Windows DPAPI).
Collector controls
You can switch off individual desktop agent collectors for a project, for example Files used in AI apps, Clipboard, Network activity or a specific tool's log reader. The process monitor that detects AI tools cannot be switched off. Hardware details and local AI apps and models are sent with the check-in and are not a collector you can switch off.
Change them in Settings > Deployment & Keys > Enroll devices. A collector
you switch off stops on each device at the desktop agent's next check-in,
about every hour, with no reinstall and no need to re-run the enrollment
script. This needs macOS agent 0.1.30 or Windows agent 0.1.21 or later. Older
agents follow the list only from the enrollment script or MDM profile. The
panel shows how many computers have confirmed the current list. The enrollment
script and MDM profile also write the list into managed settings for newly
enrolled devices: the disabledCollectors key in the
ai.agenticants.agent managed preferences domain on macOS, and
DisabledCollectors under HKLM\SOFTWARE\Policies\AntsAgent on Windows. See
Collector settings.
Employee monitoring notice
You can show the people you monitor a notice in your own words and record that each of them accepted it. The notice is off by default.
Write the notice
Go to Settings > Discover AI > Employee monitoring notice and enter the Notice text, up to 4,000 characters.
Turn it on
Turn on Show the notice to employees and save. Editing the text asks everyone to accept it again.
Where it appears. The notice is shown the first time a client runs for a user, and again whenever you change the text:
| Client | Minimum version | How it shows the notice | Picks up changes |
|---|---|---|---|
| Browser extension | 0.1.27 | Opens the notice in a new tab | At its next registry refresh, about every 10 minutes |
| macOS agent | 0.1.32 | Opens the notice in a window | At its next check-in, about every hour |
| Windows agent | 0.1.22 | Shows the notice from the tray app | At its next check-in, about every hour |
Older clients do not show the notice.
What is recorded. Each acceptance is stored for the user and device, with the notice version and the time. Acceptance records are append-only: they cannot be edited or removed from the console.
Reporting. Who has accepted the notice, on the same settings page, lists the people enrolled in Discover AI with their devices, a status of Accepted, Some devices or Not accepted yet, and when they last accepted. Search by person, filter by status, and select Export CSV to download it.
The notice is an organization-wide setting. Changing it needs Discover AI configure write access at the organization level, which Owner and Department Admin have by default. A role granted only on a project sees the settings as view only.
Directory data
If you connect Directory Integration, ANTS Platform reads department, job title, manager, name, email, employee ID and account status from Okta or Entra ID. With either connected, each Discover AI event and trace is stamped with the person's Okta user ID or Entra object ID when it arrives. See Identity provider user ID on records. Credentials are encrypted at rest.
The directory sync only reads your directory. The one change ANTS Platform can make is optional: when AI access requests grant access on approval, ANTS adds the approved person to the group you mapped to that tool. This needs a credential that can manage group membership.
Authentication and keys
Clients authenticate with your project's keys or, with device certificates, with a certificate per device. Keys can be rotated with a 7-day grace period or revoked at once. Device certificates can be revoked per device, and a project can require hardware-backed keys. See Deploying Agents & Extensions and Devices.
Auditability
-
Enrollment script downloads and device actions (revoke, block, quarantine, delete) are recorded in the audit log, with the reason entered. So are certificates issued on a software key and changes to the hardware-backed key setting.
-
Monitoring notice acceptances are kept as append-only records per user and device. See Employee monitoring notice.
-
Each client's check-ins, health and permission status are visible under Discover AI > Devices.