Skip to content

Security: qiz029/dscode

Security

SECURITY.md

Security Policy

Supported versions

Only the latest released 0.7.x version receives security fixes. Update with dscode update before reporting.

Reporting a vulnerability

Use GitHub's private Report a vulnerability form. Please do not open a public issue, and never paste credentials, session IDs or session contents into a public issue or discussion.

Include the DSCODE version (the TUI header, or npm view @toddzheng024/dscode version), your macOS version, node -v, and the smallest reproduction you can manage.

What runs where

DSCODE is a local harness: the agent loop, shell execution, file access, approvals and credentials all run on your machine.

  • Sessions, review records and cost data live in the local data directory (~/.local/share/dscode-hub, overridable with DSCODE_HOME). They are not uploaded.
  • Credentials live in owner-only local files (~/.dscode/credentials.yaml, and the DSH home's .credentials.yaml), or in the macOS Keychain for publishing tokens.
  • Cross-session messaging uses a local Unix socket and a local SQLite mailbox with owner-only permissions, reachable only by processes on the same machine.
  • The Hub client sends optional aggregate telemetry. Payloads contain no account, machine ID, IP-address field, local path, profile contents, configuration value, environment value or secret. dsh-hub telemetry off, DSH_HUB_TELEMETRY=0 or DO_NOT_TRACK=1 disable it, and DSH_HUB_TELEMETRY_DEBUG=1 prints the exact next event without sending it. See the privacy notice.

In scope

  • Sandbox or approval bypasses: an action that executes outside the configured policy.
  • Credential exposure: secrets written to logs, session events, telemetry, or files another user can read.
  • Local IPC weaknesses in the session bridge: authentication, path permissions, mailbox isolation.
  • Skill, hook or MCP handling that lets untrusted content escalate permissions.

Out of scope

  • Prompt injection and model output quality.
  • A machine that is already compromised, or a malicious browser extension.
  • Third-party MCP servers, skills or hooks you install.
  • The upstream DSH packages' own behaviour.

Hardening notes

  • Automatic review is a convenience, not a security boundary. Keep the approval policy at ask for work you do not trust.
  • Skills, hooks and MCP servers run with your user's privileges. Review what you install.

There aren't any published security advisories