Skip to main content

Swif ADR Sensor Policy

Overview

The Swif ADR Sensor Policy helps organizations observe how supported AI coding agents are used on managed macOS, Windows, and Linux devices. It uses the open-source Uber Agentic Detection and Response (ADR) Sensor to read supported local agent logs, normalize the activity into a common format, and optionally upload the resulting ADR artifacts to Swif.

ADR stands for Agentic Detection and Response. Its Sensor component is designed to provide security observability for AI agents by recording their intent, conversations, tool use, and execution traces in a consistent schema.

Use this policy to:

  • Collect telemetry from selected AI coding tools on macOS, Windows, and Linux

  • Identify which models, tools, and local agent workflows are being used

  • Support investigation of suspicious tool calls, prompt injection, unauthorized file access, and possible data exfiltration

  • Control how much recent session history is included

  • Choose how frequently the sensor runs

  • Decide whether generated ADR artifacts are uploaded to Swif

Privacy notice: ADR telemetry can contain conversation content, project paths, tool arguments, tool results, usernames, hostnames, and other sensitive data. Review the Data Collected by ADR Sensor and Privacy and Security Considerations sections before enabling this policy.

Requirements

Requirement

Details

Supported operating systems

macOS, Windows, and Linux

Device ownership

Company-owned and BYOD

Device management

The device must be enrolled in Swif Agentic Security

Swif agent

The Swif agent must be installed and operational

Local data

At least one supported AI agent must have readable local session or audit data

Connectivity

The device must be online to receive the policy and upload artifacts

Source availability varies by operating system. Claude Desktop and Warp are supported by ADR Sensor on macOS and Windows, but not Linux. Sources that are unavailable on the current operating system are skipped.

How the Policy Works

When the policy is enabled, the Swif agent runs ADR Sensor at the configured interval. ADR Sensor reads the local records created by the selected AI agents and converts them into normalized session artifacts.

The workflow is:

  1. A supported AI agent writes session, conversation, or tool-use data to its local storage.

  2. ADR Sensor reads the applicable local data source.

  3. The sensor normalizes the activity into ADR events.

  4. The configured history window determines how much recent session history is included.

  5. If Upload Enabled is turned on, Swif uploads the generated ADR artifacts for centralized security visibility.

The policy is an observability control. It does not itself prevent an AI agent from running a command, accessing a file, or using a tool. Use it with appropriate application controls, AI-agent runtime policies, endpoint protections, and access restrictions.

Default Configuration

Setting

Default

Enabled

Disabled

Sources

Empty, meaning all sources supported by this Swif policy

History Window Days

14

Collection Interval Min

60

Upload Enabled

Enabled

Because Enabled defaults to disabled, the policy does not collect telemetry until an administrator turns it on.

Important: If you enable the sensor without changing the other defaults, it scans all supported sources available to this Swif policy, includes up to 14 days of session history, runs every 60 minutes, and uploads the generated artifacts to Swif.

Supported Sources

The Sources field accepts the following values:

Source value

AI agent

Supported operating systems

Typical upstream data source

claude

Claude Code

macOS, Windows, and Linux

JSONL session files under the user's Claude project data

codex

OpenAI Codex CLI

macOS, Windows, and Linux

JSONL session files under the user's Codex data

claude_desktop

Claude Desktop local agent mode, including Claude Cowork activity

macOS and Windows

Local audit JSONL files

cline

Cline

macOS, Windows, and Linux

Local JSON task files

cursor

Cursor IDE

macOS, Windows, and Linux

Local SQLite application state

warp

Warp Terminal

macOS and Windows

Local SQLite application data

Leave Sources empty to collect from every source recognized by this Swif policy that is available on the device's operating system. To reduce collection scope, enter only the source values your organization has approved.

Source names must match the values in the table exactly. For example:

claude 
codex
cursor

The upstream ADR project may add or support other parsers. A source is not available through this policy unless it appears in the Swif policy's supported values. For example, current upstream ADR documentation lists opencode, but the current Swif ADR Sensor Policy does not expose opencode as an allowed source.

For the current upstream list and storage formats, see the ADR Sensor documentation.

Data Collected by ADR Sensor

The exact fields depend on the AI agent, its version, the available local logs, and the source parser. ADR's normalized event schema can include:

Data category

Examples

Device and user context

Hostname and local username

Session metadata

Session identifier, timestamp, source, and sequence information

Model information

Model name or identifier recorded by the source application

Project context

Local project or workspace path

Conversation content

User prompts and assistant responses

Tool activity

Tool name, tool type, arguments, results, and status

Agent configuration

Available tools, MCP servers, plugins, skills, permission mode, or other session context when the source records it

Background-agent context

Whether certain Claude Desktop sessions were interactive or dispatched

Some sources expose more information than others. If a source application does not record a field, ADR Sensor cannot collect it.

ADR Sensor reads existing local agent records and normalizes them. It does not reconstruct activity that the source application never logged.

Create the Policy

  1. Sign in to the Swif Admin Dashboard.

  2. Go to Agentic Security > Policies.

  3. Create a new policy.

  4. Select Swif ADR Sensor Policy.

  5. Enter a descriptive policy name, such as ADR Sensor - Pilot.

  6. Configure the sensor settings.

  7. Assign the policy to a small group of test devices representing each operating system in your fleet.

  8. Allow the devices to check in and apply the policy.

  9. Generate representative activity in an approved AI agent.

  10. Review the policy report and uploaded ADR data.

  11. Expand the deployment only after security, privacy, and legal review is complete.


Policy Settings

Enabled

Controls whether ADR Sensor collection runs on the device.

Value

Behavior

Enabled

The Swif agent runs ADR Sensor according to the configured interval.

Disabled

Scheduled ADR Sensor collection is turned off.

Default:

Disabled

Disabling the policy stops future scheduled collection. It does not necessarily delete artifacts that were previously generated locally or uploaded to Swif.

Sources

Limits collection to specific supported AI agents.

An empty list means all sources supported by the Swif policy. This is broad collection, not a setting that disables source collection.

For least-privilege deployment, explicitly select only the AI agents approved for the targeted users. This also reduces processing time and avoids collecting unrelated AI sessions.

Supported values:

claude 
codex
claude_desktop
cline
cursor
warp

History Window Days

Controls how much recent session history is included in generated and uploaded ADR artifacts.

Property

Value

Default

14 days

Minimum

1 day

Maximum

30 days

A shorter history window reduces the amount of historical conversation and tool activity processed during each collection. A longer window provides more investigative context but increases data volume and privacy exposure.

This setting controls the session-history window included by the sensor. Do not assume it automatically deletes previously uploaded artifacts after the same number of days. Apply your organization's retention requirements separately to data already stored in Swif or another connected system.

Collection Interval Min

Controls how often the Swif agent runs ADR Sensor.

Property

Value

Default

60 minutes

Minimum

15 minutes

Maximum

1440 minutes

Examples:

Value

Schedule

15

Every 15 minutes

60

Every hour

240

Every 4 hours

1440

Once every 24 hours

A shorter interval provides faster visibility but increases local processing, artifact generation, network use, and upload frequency. Start with 60 minutes unless your monitoring requirements justify more frequent collection.

Upload Enabled

Controls whether generated ADR artifacts are uploaded to Swif.

Value

Behavior

Enabled

Generated ADR artifacts are uploaded to Swif when the device is online.

Disabled

The sensor can run, but this policy does not upload the generated artifacts to Swif.

Default:

Enabled

Disable upload during a local validation phase if your organization must first review the content and sensitivity of generated artifacts. When upload is disabled, centralized Swif visibility will not receive those artifacts through this policy.

An offline device cannot upload artifacts until connectivity is available and the Swif agent retries or performs a later collection. The exact handling of locally generated artifacts during offline periods can depend on the installed Swif agent version.

Recommended Configuration

For an initial company-owned device pilot:

Setting

Recommended value

Enabled

Enabled

Sources

Explicitly list only approved and actively used AI agents

History Window Days

7

Collection Interval Min

60

Upload Enabled

Enabled after privacy review

Using an explicit source list and a seven-day window reduces collection scope while providing enough recent activity for validation and investigations.

For BYOD devices, do not automatically reuse the company-owned configuration. Determine whether collecting personal account conversations, personal project paths, or other non-work activity is technically avoidable and legally permitted.

Consider leaving the policy unassigned to BYOD devices unless there is a clearly defined business scope, employee notice, and approved data-handling process.


Configuration Examples

Monitor Only Approved Coding Agents

Use this configuration when the organization permits Claude Code, Codex CLI, and Cursor:

Enabled: Enabled 
Sources: claude, codex, cursor
History Window Days: 7
Collection Interval Min: 60
Upload Enabled: Enabled

The sensor ignores the other source adapters exposed by this policy.

Validate Locally Before Uploading

Use this configuration during a privacy or technical evaluation:

Enabled: Enabled 
Sources: codex
History Window Days: 1
Collection Interval Min: 60
Upload Enabled: Disabled

This configuration limits the scan to recent Codex activity and prevents this policy from uploading the generated artifacts. Centralized telemetry will not be available in Swif until upload is enabled.

Increase Collection Frequency for an Investigation

Enabled: Enabled 
Sources: claude, claude_desktop, cursor
History Window Days: 14
Collection Interval Min: 15
Upload Enabled: Enabled

Use a 15-minute interval only when faster telemetry is necessary. Return to a longer interval after the investigation to reduce endpoint and network activity.

Collect From All Supported Sources

Enabled: Enabled 
Sources: Leave empty
History Window Days: 14
Collection Interval Min: 60
Upload Enabled: Enabled

Leaving Sources empty instructs the policy to use all six source adapters it currently exposes. Review this scope carefully before deployment.

Verify the Policy

After assigning the policy:

  1. Confirm that the device is online in Swif.

  2. Confirm that the policy reports as installed or applied.

  3. Confirm that the Swif agent is running.

  4. Confirm that at least one configured AI agent has recent local session activity.

  5. Wait for the configured collection interval, or use the available Swif policy action to trigger a new check-in if applicable.

  6. Review the policy execution report for collection or parser errors.

  7. If Upload Enabled is on, confirm that recent ADR telemetry appears in the applicable Swif security view.

  8. Compare the source, timestamp, model, and tool activity with a known test session.

  9. Confirm that agents excluded from Sources do not appear in newly collected artifacts.

  10. Confirm that sessions older than History Window Days are not newly included by the sensor.

Use a controlled test session that contains no credentials, customer data, or proprietary source code. A successful pilot should prove collection behavior without intentionally exposing sensitive information.

Security Use Cases

ADR telemetry can support investigations into:

  • Unexpected or unauthorized AI coding-agent use

  • Suspicious shell commands or file operations

  • Access to sensitive paths or credentials

  • Unapproved MCP tools or servers recorded in session context

  • Prompt-injection attempts appearing in agent conversations

  • Abnormal bursts of agent or tool activity

  • Background or dispatched agents performing unattended work

  • Possible data movement or exfiltration through agent tools

Detection quality depends on the source application's local records, the configured history window, collection timing, and the analysis applied to the resulting artifacts. The sensor does not guarantee that every action is logged or that every harmful event is detected.

Uber describes ADR as combining discovery, observability, benchmarking, detection, and prevention capabilities. The Swif policy described here specifically configures ADR Sensor observability on macOS, Windows, and Linux. It should not be represented as enabling every component of the broader ADR project.

Privacy and Security Considerations

Before deployment:

  • Inform affected users that supported AI-agent session data may be collected.

  • Document the business purpose and lawful basis for collection.

  • Review whether prompts, responses, tool results, or project paths can contain secrets, personal data, health data, customer data, or proprietary code.

  • Limit Sources to approved business applications.

  • Use the shortest History Window Days value that meets the monitoring requirement.

  • Restrict access to uploaded ADR artifacts in Swif.

  • Define retention and deletion requirements for uploaded data.

  • Test BYOD deployment separately from company-owned devices.

  • Avoid using a test prompt containing real credentials or regulated information.

  • Coordinate with legal, privacy, security, and employee-relations teams when required.

ADR Sensor reads data already stored by supported applications. The policy does not change the privacy or logging settings of those source applications. Review each application's own data storage, cloud processing, and retention behavior separately.

The upstream ADR Sensor project is open source under the Apache License 2.0. Review the source code and release history as part of your software supply-chain process before production use.


Troubleshooting

The Policy Reports Installed but No Telemetry Is Collected

Check the following:

  • Enabled is turned on.

  • The device is running macOS, Windows, or Linux.

  • The Swif agent is installed and operational.

  • The device checked in after the policy was assigned.

  • At least one selected AI agent is installed and has recent local session data.

  • The source value uses the exact supported spelling.

  • The local user account that created the AI sessions is available to the sensor runtime.

  • The configured history window includes the test session.

  • The collection interval has elapsed.

Review the policy execution report for source-specific parser errors.

Only Some AI Agents Appear

Confirm that the missing agent is included in Sources, or that Sources is empty. Then confirm that the source application has produced local records in a format supported by the installed ADR Sensor version.

Supported upstream platforms do not automatically mean that every source is enabled in the Swif policy. Use only the six source values documented in this article.

Older Sessions Are Missing

Increase History Window Days if the session is older than the configured window. The maximum supported value is 30 days.

Also confirm that the source application still retains its local session data. ADR Sensor cannot ingest a session after its original records have been deleted by the application or user.

Data Is Collected but Does Not Appear in Swif

Confirm that:

  • Upload Enabled is turned on.

  • The device is online.

  • The Swif agent can reach required Swif services.

  • The collection run completed successfully.

  • The administrator is reviewing the correct device, time range, and security view.

If the device was offline, allow it to reconnect and complete another collection cycle.

Collection Uses More Resources Than Expected

  • Increase Collection Interval Min.

  • Reduce History Window Days.

  • Select only the required Sources.

  • Check whether a source contains unusually large session databases or histories.

  • Review the execution report for a parser that repeatedly fails or retries.

Do not set the collection interval below the supported minimum of 15 minutes.

An Unsupported Source Is Rejected

The Swif policy validates source names. Current accepted values are:

claude 
codex
claude_desktop
cline
cursor
warp

An upstream ADR parser is not automatically available through Swif. Contact Swif Support if your organization needs another source added.

Disabling the Policy Does Not Remove Existing Data

Disabling Enabled stops future scheduled sensor collection. It does not necessarily delete:

  • Source logs maintained by the AI application

  • ADR artifacts previously generated on the device

  • ADR artifacts already uploaded to Swif

  • Data exported or forwarded to another security system

Use the applicable retention and deletion procedures for each storage location.

Change or Remove the Policy

To change collection behavior:

  1. Edit the policy in Swif.

  2. Update the sources, history window, interval, or upload setting.

  3. Save the policy.

  4. Allow assigned devices to check in.

  5. Verify the next collection cycle.

To stop future ADR Sensor collection:

  1. Set Enabled to disabled, or remove the policy assignment.

  2. Allow the device to check in.

  3. Confirm that no new collection runs appear after the expected interval.

  4. Follow your organization's deletion process for previously generated or uploaded artifacts when required.

Removing the policy does not delete the original session data stored by Claude, Codex, Cursor, or another source application.

Notes

  • The policy supports macOS, Windows, and Linux.

  • Company-owned and BYOD devices are supported.

  • Claude Desktop and Warp sources are unavailable on Linux.

  • Sensor collection is disabled by default.

  • Upload is enabled by default and takes effect when sensor collection is enabled.

  • An empty source list means all sources supported by the Swif policy.

  • The current policy supports claude, codex, claude_desktop, cline, cursor, and warp.

  • The history window defaults to 14 days and supports 1 to 30 days.

  • The collection interval defaults to 60 minutes and supports 15 to 1,440 minutes.

  • ADR telemetry can include conversation and tool content, not only application inventory.

  • This policy provides observability. It does not independently block unsafe AI-agent actions.

Summary

The Swif ADR Sensor Policy provides centralized configuration for collecting AI-agent telemetry from supported applications on managed macOS, Windows, and Linux devices.

For a responsible deployment:

  1. Complete privacy and security review before enabling the sensor.

  2. Start with a small group of company-owned devices representing each supported operating system.

  3. Select only approved AI-agent sources.

  4. Use a short history window during the pilot.

  5. Confirm whether artifact upload is appropriate before leaving it enabled.

  6. Validate the data with controlled, non-sensitive test sessions.

  7. Combine ADR observability with access controls, prevention policies, and security monitoring.

Related Resources

Did this answer your question?