agent-privacy-checklisted

# Agent Privacy Check Give a non-specialist an evidence-backed answer about an agent's privacy boundary. Separate what the agent can do from what it actually did, and make the remaining unknowns visible. ## Scope gate Start by naming: - the agent and surface being checked, such as Codex Desktop, Codex CLI, Claude Code, or a generic agent; - the exact scope, such as the current project, one Skill or plugin, one session, or a selected configuration directory; - the trust boundary, including local files, connected services, the model provider, and external websites or APIs. Use the narrowest scope that answers the question. The default is the current project plus the agent configuration and tool metadata that are directly available to the current run. Expand to a home directory, browser profile, session archive, or external service only when the user names that scope. Use these evidence labels throughout the audit: | Label | Meaning | | --- | --- | | Observed | The current read-only check directly saw this capability or action. | | Available | Configuration or permissions make it possible, but this run did not prove that it happened. | | Not found | The named scope was inspected and no matching evidence was found. | | Unknown | The scope or runtime boundary could not be inspected reliably. | Treat a declaration in a Skill, README, prompt, or tool description as a lead, not proof of runtime permission. Look for the permission, implementation, configuration, or activ