# Security overview Source: https://docs.ideharmony.com/security This page is written to be useful during an incident, which means it is more interested in the boundaries than in reassurance. ## What Harmony is A local-first application (**Harmony Core**). The daemon, terminals, tasks, agent-to-agent messaging, voice models, and MCP integration run on your machine. Local use does not require an account. Optional [Harmony Pro](/docs/cloud) sends only files you confirm; it is not on the path of Core work. ## What Harmony protects | Protection | Detail | |---|---| | Download integrity | Installers and updates verify exact byte size and SHA-256 before executing | | Release honesty | Unsigned builds can reach the preview channel only, and are labelled | | Approval gates | Actions that change your machine or repository can require explicit consent. Agents propose; you dispose. | | Credential isolation for managed harnesses | User-managed external harnesses do not share credentials with other providers | | Cloud transfer path blocks | Known credential-looking paths are refused on send-to-cloud — not a complete secret scanner | | Remote message containment | LAN Link messages render as text only; never pasted into a PTY, never routed to an execute path | | Schema-validated IPC | Inter-process messages are validated rather than trusted | ## What Harmony does not protect Being explicit here matters more than sounding secure. > [!SECURITY] > **Extensions, skills, and plugins are not sandboxed by default.** Third-party > extension code runs with the privileges of the process that loaded it. Do not > install extension code you would not run directly. | Not protected | Why it matters | |---|---| | Agent actions within a worktree | Worktrees are file isolation, not a security boundary | | Your provider credentials | Core does not store them; the provider CLI does. Cloud may store *other* user-configured secrets you add — [credentials](/docs/security/credentials) | | Your shell environment | Keys exported globally are visible to everything you run | | Every secret in a Cloud upload | Only known path patterns are refused; review the file list | | Content produced by a model | Treated as untrusted input, not as instruction | | Repository content | A repository can contain hostile text aimed at an agent | | Browser content | Anything a browser surface loads is untrusted | ## Untrusted inputs Treat all of these as attacker-controlled: - Model output - Repository file contents - Web pages loaded in a browser surface - Extension and skill code - Provider output - Messages received from peer machines The practical consequence: an approval prompt is not a formality. It is the last point where a bad instruction that entered through any of the above is cheap to stop. Harmony does not apply changes as a perfect unattended process. ## Windows signing status The Windows preview build is **not Authenticode signed**. SmartScreen will warn, and that warning is correct. > [!SECURITY] > Verify the published SHA-256 yourself rather than disabling SmartScreen. > Turning off a machine-wide protection to install one application leaves the > machine weaker permanently. ## Reporting a vulnerability Report security issues privately rather than in a public tracker, and include enough detail to reproduce. Do not include credentials or customer data in the report. ## Related - [Privacy](/docs/security/privacy) - [Credentials](/docs/security/credentials) - [Permissions and approvals](/docs/concepts/permissions-and-approvals) - [Known limitations](/docs/reference/known-limitations) - [Choose your setup](/docs/getting-started/choose-your-setup)