TermPlete handles terminal commands, captured output, model requests and provider credentials. Treat accidental disclosure of these data, bypasses of filtering, and unintended command execution as security issues.

Once the project is hosted on GitHub, use Security → Report a vulnerability if private vulnerability reporting is enabled. If that option is unavailable, ask the maintainer for a private reporting channel without posting exploit details or sensitive data in a public issue. No dedicated security email or response-time guarantee has been established yet.

Provide the affected version or commit, platform and shell, the relevant feature, and a minimal reproduction using synthetic values. Explain the expected and observed boundary. Do not send real API keys, credential files, full shell history or private terminal recordings.

If a key may have leaked, revoke or rotate it with the provider first. Removing the local keychain entry does not revoke the remote credential.

• API keys live in the OS credential store, with an optional environment fallback.
• Commands and output are scrubbed before storage and outbound use. Recognized credential paths and configured private paths stay out of model requests.
• A local review helper can receive hard-filtered soft private metadata only through a loopback endpoint. A remote generator receives fully filtered data.
• Explicit privacy denials and failed required local review prevent generation.
• Hybrid fallback under Block sends only a deterministically checked active prompt and omits all remembered context.
• Terminal output and model replies are untrusted. Accepting a suggestion changes the prompt; execution requires a separate user action.
• ChatGPT completions disable inherited Codex apps, plugins, MCP servers and lifecycle hooks and notification commands, and withhold terminal context until an ephemeral thread and disabled integrations are confirmed.
• Output-capture framing uses a private per-session token. Configuration and default socket directories are checked for ownership and private permissions.

These measures reduce accidental exposure. They do not guarantee recognition of every secret, correctness of a model decision, or safety of a suggested command. A loopback model may itself have logging or network access. The terminal integration is not a sandbox against programs running as the same user.

For storage controls and removal, see the usage guide. Release archives are verified against published SHA-256 checksums; this checks integrity relative to the manifest and is not an independent signature or attestation. Review the release runbook for the distribution contract.

The project is in early development. There is no long-term support branch or published security-support window. Reports should identify the exact version or commit. Maintainership and release availability are described in the repository documentation; release automation does not mean a particular version has been published.