Mindlas runs as Claude Code hooks on your machine. By design it makes no network calls and no LLM calls: every signal is computed locally from the per-session event ledger. It writes only to ~/.mindlas/ (per-machine session ledgers), the project-local .mindlas/ directory (packs, reports, stop boundaries), and — when installing — your Claude Code settings.json (reversible via --uninstall). The trust gate and loop guard can block a tool call only in their explicit opt-in modes (MINDLAS_GATE=live, MINDLAS_LOOP_GUARD=block); the default posture is observe/warn.

Open a private security advisory on GitHub (https://github.com/Evolutionairy-AI/MINDLAS/security/advisories) or contact the maintainers through the repository. Please do not file public issues for exploitable problems.