2026-10-12 Changes confirmed medium confidence

GitHub Adds Identity and Content Controls to Copilot Telemetry

Enterprise administrators can now separate user and device attribution from five optional classes of sensitive Copilot session content.

Edited by Tyronne Panaino

GitHub expanded its Copilot client telemetry controls on October 12, letting enterprise administrators attach user and device identity to agent activity while keeping prompts, responses and tool content outside the export unless those classes are separately enabled. The official GitHub changelog says every new capture class is off by default.

The change affects teams using OpenTelemetry to investigate Copilot activity across the Copilot CLI, the GitHub Copilot app and Visual Studio Code sessions that use Agent Host. Its practical value is finer separation between who or which machine produced an event and the potentially sensitive material inside an AI session. It does not establish that an exported trace is a complete audit record.

Identity can be enabled without conversation content

The new `telemetry.capture.identity` setting adds three attributes. `user.name` carries the authenticated Copilot account name, currently the GitHub login. `process.user.name` identifies the local operating-system account, and `host.name` records the local machine name.

GitHub says identity capture is disabled by default and remains separate from its existing pseudonymous user identifier. Administrators can therefore choose explicit attribution without also turning on prompts, model responses or tool content. That boundary is important for organizations that need to investigate activity but do not want every trace to contain the substance of a developer's conversation or workspace.

The three identity fields still require governance. A GitHub login, operating-system account and machine name can connect an event to a person or device even when the prompt itself is absent. The announcement defines the fields and defaults, but it does not provide a retention schedule, access-control model or independent privacy assessment for a customer's chosen telemetry destination.

Five content classes become separate choices

Administrators can independently control five additional classes: prompts, responses, tool arguments, tool output and policy details. GitHub says prompt capture may include user, system and developer prompts plus supported input context. Response capture may include assistant output and supported reasoning content.

Tool-argument capture can include raw arguments, commands, schemas and identifiers for tools, MCP servers, plugins and hooks. Tool-output capture can include results and execution-error details. Policy-detail capture can expose raw policy information such as sandbox filesystem paths. All five classes are off by default.

This turns content collection into a set of discrete decisions rather than one undifferentiated switch. It also makes the review burden more specific: source code, file contents, commands, tool results and workspace or user-profile paths can enter telemetry when the relevant classes are enabled. Teams need to align each class with the purpose of the audit and the controls on the receiving endpoint.

Managed policy overrides local configuration

GitHub says explicit enterprise-managed values take precedence over local settings, SDK configuration and environment variables. That gives administrators a centrally enforced starting point across the supported clients instead of relying on every developer to preserve the same configuration.

The older `captureContent` setting remains available as a shorthand for the five content classes, but GitHub says it never enables identity. That distinction helps separate older all-content configurations from the new identity switch, yet it also means administrators should inventory both the legacy shorthand and the granular keys when reviewing an existing deployment.

The update adds configuration capability, not proof of audit completeness. The fetched announcement does not quantify event loss, export latency, performance overhead or field consistency across the three supported client surfaces. It also does not show that telemetry prevents an unsafe agent action; the controls describe what may be recorded after or during activity, not an enforcement mechanism.

Evidence quality and next checks

The release details, default-off behavior and field definitions come from GitHub's first-party changelog. Internal confidence is medium because no independent test or fuller operational evaluation was fetched. Organizations should verify the actual attributes delivered by each client, confirm that their collector and storage rules match the selected classes, and test whether the resulting traces answer their investigation needs without collecting unnecessary content.

The next useful evidence will be detailed administration documentation, client-version coverage and measured behavior under real exports. Until then, the supported conclusion is narrow: GitHub has added independent identity and content-capture controls to Copilot telemetry, with every new class disabled by default.

Status

Confirmed. GitHub documents the October 12 configuration change and its supported client surfaces; trace completeness, operational cost and privacy outcomes remain unverified.