Logica Voice is in early development (v0.x). Only the latest minor version receives security fixes.

When v1.0 ships, we'll publish a longer-term support policy.

Please do NOT open a public issue for security vulnerabilities.

Email: sirambrosio@gmail.com with subject [SECURITY] Logica Voice — <short description>

Include:

• Description of the vulnerability
• Steps to reproduce (or proof-of-concept code)
• Affected version(s)
• Suggested fix (optional)
• Your contact info for follow-up

• Within 48 hours: acknowledgment that we received your report
• Within 7 days: initial assessment and severity classification
• Within 30 days: fix released (for critical/high severity) or detailed mitigation plan

We'll credit you in the release notes unless you prefer anonymity.

These parts of Logica Voice handle sensitive data — please pay extra attention:

• WhatsApp (@logica-voice/adapter-whatsapp): handles auth credentials in authDir. ACL is fail-closed by design — bypassing it is a security bug.
• Telegram (@logica-voice/adapter-telegram): bot token in env. ACL check on allowedChatIds.

• Voice embeddings (config/voices/*.npy): can be used to clone voices of real people. The framework does NOT verify consent — that's the user's responsibility.
• Reporting concerns: if you find a way the system enables non-consensual cloning at scale (e.g., bypassing the warning prompt), report it as a security issue.

• API keys stored in logica-voice.yaml or .env should never be committed. Our .gitignore covers .env* by default.
• LogicaOS bridge: validates response from /api/chat/stream but trusts the brain's tool calls. If you find a way for a malicious response to escalate privileges, report it.

• pgvector / Chroma: stored conversation history may contain PII. Apps using Logica Voice should comply with GDPR/LGPD.

These are documented design choices, not bugs:

• WhatsApp via Baileys runs on a gray-area WhatsApp ToS. Use at your own risk.
• No end-to-end encryption between Logica Voice server and channel adapters (they share a process). E2E would require restructuring.
• Brain adapters trust their endpoints — a malicious OpenAI-compatible server could inject content. Use trusted providers.

1. Never commit .env — use .env.example as template
2. Use ACL fail-closed — empty allowedPhones/allowedChatIds = deny all (default)
3. Rotate API keys regularly
4. Run on isolated user account — don't give Logica Voice root
5. Keep dependencies updated — pnpm update periodically
6. Backup authDir (WhatsApp) — losing it requires re-scanning QR

Reports made under this policy are not subject to additional legal restrictions beyond the MIT license. Reporters retain authorship.