Report vulnerabilities privately to the Hasna maintainers. Do not file public issues for exploitable command execution, credential exposure, prompt injection, provider routing, or MCP boundary problems.

tai is designed to preview commands and classify risk before execution. Treat command classification as defense in depth, not as a sandbox. Users remain responsible for confirming commands before they run.