Docs/Shadow Ai/Access Policy

Access Policy

The access policy decides, for each person, device and AI tool, whether the tool is allowed, shows a warning, asks for a reason, or is blocked. One policy applies to your whole organization. It replaces the earlier Warn mode setting.

The platform resolves the policy on the server and sends each device only the answer for each tool. Devices never receive your rules, your groups or anything about other people.

⚠

Endpoint versions. The policy is part of ANTS Platform v0.1.400. It is enforced by:

  • Browsers: the next browser extension release (after 0.1.30), on Chrome, Edge and Firefox.
  • Macs: macOS agent 0.1.34 or later, for desktop AI apps, command-line tools, Safari and browsers without the extension.

Neither release is available yet. Until they ship, released browser extensions (0.1.23 or later) show the existing Warn banner on every tool the policy does not allow, and nothing is blocked. Windows has no enforcement. See Older clients.

How a decision is made

For each tool, the platform checks your rules from top to bottom:

  1. A rule that is turned off is skipped.
  2. The first rule in Enforce mode that matches decides the action.
  3. If no rule matches, the default for the tool's status applies.

A matching rule in Test mode never acts. It records what it would have done, and the check continues with the next rule. See Test mode.

Two things always apply on top of the result:

  • While the Block switch is off, Block is delivered as Warn.
  • ANTS Platform hosts are never restricted, so a rule can never block the ANTS request page.

Actions

ActionWhat the person gets
AllowThe tool works as normal. Use is still recorded.
WarnA warning shows first. The person can continue.
Warn with reasonThe person must type a reason before continuing.
BlockThe tool does not open. The person sees the block page.

Before you start

  • Permissions. Viewing the policy needs read access to Configure (policy, registry, keys, directory). Changing it needs write access to that module at the organization level. Owner and Department Admin have it by default. Without it the page shows "View only. Needs organization Discover AI configure access." See Permissions.
  • Device certificates for people and device rules. The platform knows the person and the device only when the client signs in with a device certificate. With the shared API keys, only rules for Everyone and the defaults apply, and the device counts as unmanaged, with no agent, no Falcon, no extension and an unknown platform.
  • Directory for group, department and job title rules. People are matched to your directory by email, then by user principal name. Only active directory users match.
  • Macs: the Network Extension permission granted, the Network activity collector on, and an access policy signing key on your ANTS Platform deployment. Without the signing key the Mac blocks nothing. Contact your ANTS representative to confirm it is set up for your deployment.

Open the access policy

Go to Settings > Discover AI. The Access policy section sits below Sanctions. Its header shows where you stand:

StatusMeaning
Nothing live yetNo version has been published. Employees see no change.
Live: version NVersion N is what devices enforce.
Changes not live yetYou have edits that are not published.

Every edit is saved to a draft automatically. Devices keep enforcing the live version until you select Make live.

The header also has Try it out (the simulator) and a menu with Discard changes (go back to the live version) and Stop all blocking.

The policy is built in three numbered steps, followed by an Advanced section.

Step 1: Default for each tool status

The action for tools that no rule matches, by sanction status:

StatusWhich toolsStarting default
SanctionedTools your organization approved.Allow
ToleratedTools you accept but have not approved.Allow
UnsanctionedTools someone marked as not allowed.Warn
UnreviewedTools with no status yet, including every registry tool nobody has classified.Allow

Set sanction statuses under Sanctions or on the tool's panel. See Sanction status.

ℹ

In the AI Tools views and the Sanctions list, a tool nobody has reviewed is still shown as Unsanctioned. In the access policy it counts as Unreviewed, so the Unsanctioned default applies only to tools someone explicitly marked Unsanctioned.

Step 2: Rules

Rules are exceptions to the defaults. The first match wins, so put specific rules above broad ones. Select Add rule to open the rule form.

FieldWhat it does
NameRequired, up to 100 characters.
ActionAllow, Warn, Warn with reason or Block.
ModeEnforce or Test. See Test mode.
ServicesTools, Status (Sanctioned, Tolerated, Unsanctioned, Unreviewed) and Risk (Low, Medium, High, Critical). Any one of these matches. Leave all empty for every tool.
PeopleEveryone, or Selected people with an Include list of Groups, Departments, Job titles and People (email). Any one entry matches. An Exclude list with the same fields removes people from the rule.
DevicesManaged, Agent healthy, Falcon present and Extension present, each Any, Yes or No, and Platform (macOS, Windows, Linux, ChromeOS). Every condition you set must match.
WhereSurface: Browser, Mac apps or Command line. Leave empty for every place.
MessageOptional Message text, Approved alternative and Request link for this rule. An empty field uses the default from Step 3.

A rule applies only when every section matches: services, people, devices and where. A combination such as "High risk and Unsanctioned" cannot be written as one rule, because the Services fields match when any one of them matches.

People. Selected people match only people found in your directory. Exclude by email also works for people who are not in the directory; exclude by group, department or job title needs the directory. The group, department and job title lists offer the values from your directory.

Devices. Device conditions come from ANTS records, never from the device:

ConditionTrue when
ManagedThe client, or another client on the same device, has an active device certificate.
Agent healthyA desktop agent on the same device checked in within the hours set under Timing (24 by default).
Falcon presentA desktop agent on the device reported a CrowdStrike Falcon agent ID. Falcon agent ID collection is off unless it was turned on for your project.
Extension presentThe client is the browser extension, or a browser extension for the same person and computer checked in within the last 24 hours.
PlatformThe platform the client reports. A device with an unknown platform does not match a platform condition.

Where. Browser is the browser extension. Mac apps is desktop apps on a Mac, and also Safari and other browsers the Mac agent covers. Command line is command-line tools on a Mac.

The rules table lists Name, Services, People, Devices, Action, Mode and an On switch. Search by name and filter by action or mode. Drag rows to reorder them, or use the row menu: Edit, Duplicate, Move up, Move down and Delete. Reordering by drag is off while a filter is applied. A policy holds up to 200 rules.

Step 3: Messages

The default text shown to people, used by every rule that does not set its own:

FieldWhat it does
MessageUp to 600 characters.
Approved alternativeThe tool people should use instead, up to 200 characters. A full https:// link shows as an Open button on the block page.
Request linkWhere Request access goes. Must be an http or https link.

When AI Access Requests is turned on, Use ANTS request page fills in your organization's request page. The browser extension adds ?tool=<tool ID> to the link so the request form opens with the tool filled in.

Previews show the Block page, the Banner and the Mac notification.

The Block switch

The Block switch at the top of the section is off for every organization until someone turns it on.

  • While it is off, nothing is blocked: a Block result is delivered as Warn.
  • You cannot make a policy live that has an enforced Block rule or a Block default while the switch is off. Make live is refused with "Block is switched off for this organization." Put Block rules in Test mode to try them first.
  • Turning it on asks for step-up verification when that is enabled.

Test mode

A rule in Test mode never blocks or warns anyone. In the browser, when a Test rule matches and its action is not Allow, the extension records a "would have blocked" or "would have warned" event, at most once per tab, tool and rule every 10 minutes. A Test rule for Block keeps recording "would have blocked" while the Block switch is off, so you can see what Block would do before you turn it on.

Macs do not record Test mode events.

Advanced

Timing

SettingDefaultRange
Hours a reason is valid81 to 72. After a person gives a reason, they can use that tool for this many hours.
Agent healthy after check-in (hours)241 to 720. Used by the Agent healthy device condition.

Select Save to apply. Timing changes take effect without publishing a version.

Rollout

Shown once a version is live, and refreshed every 30 seconds. It counts browser extensions and macOS agents:

CountMeaning
On this versionClients that confirmed the live version.
WaitingClients new enough to enforce that have not confirmed the live version yet.
Older softwareExtensions and agents too old to enforce. They keep the Warn banner behavior.
Not seen in 7 daysNo check-in for a week.

A Mac confirms a version only after its network extension has verified and adopted it.

History

Every published version is kept and cannot be changed. The table lists Version, Published by, When and Note. Compare lists what changed from that version to the live one: defaults, rules added, removed, changed or moved, and the default message. Restore publishes an older version's content as a new version, with the note "Restore of vN". A draft you have open is kept.


Try it out

Try it out shows what a person on a given device would get for a tool, using the same evaluation the devices get. Enter:

  • Person (email). Leave it empty for someone not in the directory.
  • Tool, Surface and Policy (Draft or Published).
  • Managed, Agent healthy, Falcon present, Extension present and Platform.

Select Run test. The result shows the action and why, for example "Rule: Block DeepSeek" or "Default for Unsanctioned tools". It also notes "Not in directory. Only rules for everyone apply.", a Block shown as Warn because the switch is off, and any Test rule that would act. Each rule is listed as matched, off, or no match with the sections that failed.

The Tool Registry opens the same panel from a tool's Effective action badge.


Make live

Select Make live

The button is available when you have changes that are not live.

Review the impact

Publish policy shows the impact over the last 7 days: how many uses by how many people Would block, Would ask for a reason and Would warn, and what Test rules would block. It counts browser prompt activity in the current project. Rules with device conditions are counted as if every device were unmanaged. Below it, the changes from the live version are listed.

Add a note and publish

Add an optional Note (up to 300 characters) and select Publish. This asks for step-up verification when it is enabled.

Publishing is refused if the live policy changed after you started your draft, or if someone else published at the same time. Reload and try again.

Stop all blocking

From the header menu, Stop all blocking publishes a new version at once (after step-up verification): every Block, Warn and Warn with reason rule moves to Test mode, and any default set to Block becomes Warn. Allow rules stay on, so your exceptions still hold. A draft you have open gets the same change. To undo it, restore the earlier version.


What employees see

In the browser

Needs the next browser extension release (after 0.1.30), signed in. The extension checks the address of every page the person opens on a registry tool or internal AI tool, with the same domain and path matching as before. When several tools match, the most specific one wins.

Block. The page is replaced before it loads by a block page titled "Tool is blocked", with your message (or "This AI tool is not approved." if the message is empty). It offers:

  • Open site, when the approved alternative is a link, or the text "Use alternative instead."
  • Request access, when a request link is set.
  • Go back.
  • Why am I seeing this, with the rule name (or "Default for this tool") and the policy version.

Warn. The banner across the top of the page, with your message, the approved alternative and the request link. The person can dismiss it and continue. It behaves as described in How the banner behaves.

Warn with reason. Before the page loads, a page titled "Reason needed for Tool" asks Why do you need this tool, with the hint "10 to 500 characters. Valid for N hours." Continue stays disabled until the reason is long enough. Continue saves the reason and opens the page the person asked for. The page also has the alternative, Request access, Go back and Why am I seeing this.

After a reason, the tool opens with no prompt and no banner for the hours a reason is valid. The allowance is kept in that browser. When the reason is recorded by the platform (device certificates only), it also applies on the person's other devices, as long as the tool still resolves to Warn with reason under the same rule.

When a new policy arrives, tabs already open on a tool that is now blocked or needs a reason are redirected straight away, and text typed in them is lost. Tell people before you first make Block live.

On a Mac

Needs macOS agent 0.1.34 or later. The agent's network extension checks each new connection against the hosts the Tool Registry lists for each tool on macOS, and their subdomains.

WhereCovered by
Desktop AI appsThe Mac agent (Mac apps)
Command-line toolsThe Mac agent (Command line)
Safari, and browsers such as Brave, Arc, Opera and VivaldiThe Mac agent (Mac apps)
Chrome, Edge and FirefoxThe browser extension when it is live for that person and computer, otherwise the Mac agent
  • Block: the connection is dropped and a notification titled "Tool is blocked" shows your message, "Use alternative instead." and "Click to request access." Clicking it opens the request link.
  • Warn and Warn with reason: the connection is allowed and a notification titled "Tool is not approved" shows the same text. There is no reason prompt on the Mac.
  • A notification shows at most once per tool and action every 15 minutes. People must allow notifications from the agent to see them.

The agent fetches its policy about every 2 minutes. The platform signs each answer for that organization and device, and the network extension enforces only a policy whose signature it can verify. It keeps the last verified policy for up to 24 hours after the platform resolved it, then stops enforcing. A Mac that has never received a policy enforces nothing.

Older clients

Extensions and agents that cannot enforce keep working:

  • Browser extensions 0.1.23 and later show the Warn banner on every tool whose result is Warn, Warn with reason or Block, with your default message and request link. If no message is set, the banner says "This AI tool is not approved for company data." If no request link is set, it links to your ANTS request page. Rule messages and Test rules do not affect it, and nothing is blocked.
  • macOS agents before 0.1.34 and all Windows agents never block or notify.

How changes reach devices

ChangeWhen devices pick it up
A published version, a restore or Stop all blockingA browser that is sending activity refetches its settings when it next uploads events. Every browser refreshes about every 10 minutes. Macs fetch the policy about every 2 minutes.
Block switch and timingThe same way.
Sanction status of a toolThe browser extension's next registry refresh, and the Mac's next policy fetch.
A person joining or leaving a directory groupAfter the next directory sync, when the person's devices next refetch.

When an approved AI access request adds someone to the group a rule allows, they get access after the next directory sync, with no policy change.


Where results appear

  • Sanctions (Settings > Discover AI): an Access column with the action for each tool.
  • Tool Registry (Settings > Tool Registry): an Effective action column. Select the badge to open Try it out for that tool. It shows what someone outside the directory gets on an unmanaged device, and Varies when a managed device would get a different result. It shows -- while nothing is live.
  • AI tool panel and user page (Discover AI): an Access card with Blocked, Warned, Reasons given and Test would block counts for the last 30 days, and the 20 most recent Reasons with the person's email and date. On a user page it also lists the tools that person is blocked from or needs a reason for, under the live policy on an unmanaged device. The card needs organization-level Configure read access.

Events, automations and SIEM

EventWhen
BlockedThe browser showed the block page, or a Mac dropped a connection.
Would block, would warnA Test rule matched in the browser.
Warning shown, dismissedThe Warn banner showed or was dismissed.
Reason givenSomeone continued with a reason.
Enforcement fallbackThe browser could not install its blocking rules, so it redirects after the page starts loading.

Each event carries the tool, the rule, the policy version and the surface.

Automation rules can trigger on AI access blocked and AI warning justified in the Discover AI group, with the tool, rule, user email and host. They fire at most once a minute per person, tool and event type. AI warning justified fires only for reasons the platform recorded, which needs device certificates. Test mode events do not trigger automations.

The SIEM export sends these events as tool activity with the activities access_blocked, access_would_block, access_would_warn, warn_justified, enforcement_fallback and native_blocked, plus the rule ID, policy version, surface and enforcement method.

Audit log

The audit log records draft saves, discarded drafts, publishes, restores and Stop all blocking as aiAccessPolicy entries (actions saveDraft, discardDraft, publish, restore and killSwitch), and changes to the Block switch and timing as aiAccessSettings entries (action update). Reasons people give are not written to the audit log.

For what is stored when someone is blocked or gives a reason, see Data & Privacy.


Permissions

ToYou need
See the access policy, Rollout and HistoryConfigure (policy, registry, keys, directory) read
Use Try it out, see the rule pickers and the Access cardConfigure read at the organization level
Edit, make live, restore, Stop all blocking, change the Block switch or timingConfigure write at the organization level

Owner, Governance Owner and Project Admin have organization-level Configure write by default. Making live, restoring, Stop all blocking and turning Block on also ask for step-up verification when it is enabled. See Members & Roles.


Coverage and limits

  • Browser: page loads in the main frame only. Embedded frames, API calls from other pages and private windows where the extension does not run are not covered. See Disable private browsing.
  • Signed out: an extension that is signed out, or waiting to re-enroll, enforces nothing.
  • Tracking switched off: a tool whose tracking is switched off for the extension with disabledObservers in its managed settings is not enforced.
  • Mac: only hosts the registry lists for a tool on macOS are checked. Shared AI provider API hosts such as api.openai.com are never blocked, so a tool that only calls a shared provider API cannot be blocked on the Mac. Connections with no host name, and very short command-line runs that end before they are checked, are allowed.
  • Mac network collector: if you switch off the Network activity collector, the agent stops handing new policies to the network extension. The extension keeps enforcing its last verified policy for up to 24 hours, with no notifications and no block events.
  • Shared API keys: person and device rules do not apply. On a Mac with API keys, the agent also covers Chrome, Edge and Firefox.
  • Windows: no enforcement.
  • Warn with reason is a speed bump, not a lock: the allowance is kept in the browser, and a person who can change the browser's stored data can extend it.

Moving from Warn mode

Warn mode is replaced by the access policy:

  • If your organization had Warn mode on, version 1 was published for you with the note "Migrated from Warn mode settings": Unsanctioned and Unreviewed set to Warn, Sanctioned and Tolerated set to Allow, and your banner text, approved alternative and request link as the default messages. Employees see the same banners as before.
  • If your organization was in Observe mode, nothing is live and employees see no change until you make a version live.
© 2026 ANTS Platform, Inc.Docs v1.0 · Last updated June 2026