repoglot is a Bash + Python harness that orchestrates external translation runs via the Claude Code CLI. It does not handle user-supplied secrets, does not run a network service, and does not process untrusted input. The realistic threat surface is small.

If you discover a security issue nonetheless — particularly something that could trick the harness into executing unexpected commands, leaking content outside the target repo, or bypassing the validator guardrails — please report it privately to christopher.louet@pm.me rather than opening a public GitHub issue.

I aim to acknowledge a security report within 7 days and to ship a fix or documented mitigation within 30 days of triage.

In scope :

• Command injection in any of the scripts/*.sh or core/validators/*.sh.
• Path traversal that lets scan-project.sh or translate-batch.sh write outside --root.
• Validator bypasses where a translated file passes check-refs.sh / check-glossary.sh / profile hooks despite materially breaking the blacklist or glossary contract.
• The install.sh bootstrap (e.g., MITM on the curl | bash path is the user's responsibility, but please report any signature / integrity improvements you'd like to see).

Out of scope :

• Vulnerabilities in upstream claude CLI, python3, bash, or system packages.
• DoS via crafted inputs that legitimately exhaust the user's Claude quota.
• Bugs in the translation quality of claude --print — those are model behaviour, not security issues.

Only the latest v1.x release. No back-ports.