# Security policy ## Supported versions During the prerelease phase, security fixes target the latest published Leogriel prerelease. Historical `@skillctl/*` packages and older Leogriel prereleases receive migration guidance but no parallel feature development. ## Reporting a vulnerability Do not disclose a suspected vulnerability in a public issue, discussion, pull request, test artifact, or behavioral-runner log. Use GitHub's private vulnerability reporting for `xFurti/leogriel` when it is available. If the repository does not expose that form, contact `@xFurti` through a private contact method listed on the maintainer's GitHub profile and include only the minimum information needed to establish a private channel. Include: - affected package and version; - operating system and Node version; - impact and required preconditions; - minimal reproduction steps; - whether credentials, untrusted plugins, test YAML, or unmanaged agent targets are involved. Never include real API keys, npm tokens, ChatGPT authentication-profile files, cookies, or private skill contents. Use synthetic values and redact paths that identify people or private repositories. ## Security boundaries - Leogriel does not use telemetry. - The default audit is static and offline. - npm packages use integrity checks and release provenance, but source trust patterns are policy signals rather than cryptographic signatures for arbitrary skills. - Plugins execute Node.js with the user's permissions and are not sandboxed. - Behavioral tests and command assertions may execute untrusted code. Network denial, isolated homes, environment filtering, timeouts, and process cleanup reduce risk but do not create an absolute sandbox. - Retained workspaces may contain sensitive files generated by an agent. - Unmanaged agent targets are never overwritten without explicit selection, confirmation, and backup. ## Coordinated handling Maintainers will validate the report privately, identify affected versions, prepare regression tests and a fix, and publish an advisory when disclosure no longer increases user risk. No fixed response deadline is promised during the prerelease phase.