RuleGuard is intended to flag policy-violating or safety-critical user/content manipulations in multi-turn agent UIs (indirect injection, credential phishing, tool-argument smuggling, privilege override framing). It is not a full replacement for human moderation, DLP, or network isolation.

• The safeguard_api → classify_mini_templar → safeguards_adapter.classify_text path is designed to run without outbound network when DARKSPACE_OFFLINE_ONLY=true (default in Dockerfile).
• Hugging Face: Model weights for sentence-transformers/all-MiniLM-L6-v2 are baked at image build time; runtime should not need registry access if the image was built successfully.

• DARKSPACE_HMAC_SECRET in Dockerfile is a placeholder for local builds. Override in production via orchestrator secrets; never commit real keys.

• Pin versions in requirements-docker.txt. Rebuild images when upgrading transitive dependencies; run pytest before release.

Use your organization’s coordinated disclosure channel for production vulnerabilities. For this OSS repository, open a private security advisory with the GitHub maintainers where available.