Policies

A policy is a trigger, a filter and a list of actions. The trigger says what kind of thing the policy is about, the filter says which of those it matches, and the actions say what happens when it does.

The four types

• Network — an interaction with an AI service, as it happens. Narrowed to All, Text or Attachment.
• Host — a file on a machine: one that is there, or one that appears, changes or goes. Narrowed to Any, or one category of AI artifact, such as an agent skill or an MCP connection.
• Hook — an action an AI tool is about to take, before it acts. Narrowed to the action: a prompt, a shell command, a file read or write, a web fetch, a tool or MCP call, or a session start.
• Connector — a record on a connected platform, on every sync. Narrowed to the kind of record.

The Policies page has a tab for Network, Host and Hook. Connector policies are on the fourth tab, From connections, grouped by connector, with the rules each connection runs. A connector rule has one switch, and it belongs to the connection.

Writing a policy

Choose New policy. The screen opens on Start from: a built-in starter such as Secrets in prompts or New MCP connection, or Start blank to choose the trigger yourself.

The page is then read top to bottom:

1. Trigger — the type, and what narrows it.
2. Filter — the rows a match must meet. Each row starts with Look at: what the row asks about. A row can read an engine's answer, such as AI judge › Skill Scanner. A policy with no filter rows applies to everything its trigger sees.
3. Actions — what happens on a match.
4. When no verdict arrives — Network only. Whether the interaction is let through or stopped when an answer does not come in time, and how long to wait.
5. Where it runs — how many machines the policy runs on, in one sentence. On a Connector policy it names the connection and when it last synced.

Beside them, Identity holds the name and description, and Applies to picks the machines: every machine, only the machines of some enrollment profiles, or every machine except those. Choose Create to save it.

Actions

What a policy can do depends on its type.

• Network: Block, Ask, Redact, Raise an alert, Notify and Open an issue.
• Host: Tag it and Open an issue. A policy set to fire when something happens to a file can also Raise an alert and Notify.
• Hook: a decision — Allow, Ask, Block or Lock — and, besides it, an alert, a notice or a webhook.
• Connector: Raise an alert, Open an issue, and the actions the connection can take on the platform, such as Revoke the session.

Four of them need a word more:

• Ask holds the action until the person at the machine confirms it. With Require a reason on, they must also type why. Nobody else approves anything; the record keeps their answer.
• Redact replaces the personal data and secrets that the filter's PII and secrets rows name before the model receives them. It is offered only when the filter has such a row.
• Notify raises the alert and also shows a notice to the person at the machine. A notice is shown once and not kept.
• Open an issue files an issue under Detections › Issues, for something that stays true until it is fixed. See Alerts and issues.

Enabled, Disabled and Monitor

A policy is always in one of three states.

• Enabled — takes its actions.
• Disabled — is on no machine.
• Monitor — records every hit and raises the policy's alerts and issues. It stops and changes nothing: no block, redaction, ask, notice, fix or action on a connected platform. It is how to see what a policy would do on real traffic before it does it.

A new policy starts disabled, on no machine. Changing the state sends the policy to machines in the same step, and each machine picks it up on its next heartbeat.

Groups, and policies Oxelone provides

A group is a named, ordered set of policies of one type, with its own On · Off. Off stops every policy in it and keeps each one's own state; On gives it back.

A policy, group or connector rule marked Provided by Oxelone comes with the product. Its rule cannot be edited in the workspace; Duplicate to edit makes a copy that can.

What a policy did

The list's Hits column counts the last seven days of matches, whatever the policy's state — so a Monitor policy shows what it would have done. A policy's page has a Rule tab and an Activity tab. Activity charts its decisions over time, lists the recent ones and opens them in Governance › Runs, where every decision of every policy is listed.