Please do not open a public issue for security reports. Email the maintainer instead (see the GitHub profile) with a description and reproduction steps. You'll get a response within 72 hours.

Consensus handles workspace conversation data, so these hold everywhere:

• Fail-closed permissions — private-channel decisions are only shown to verified channel members; membership-check failures redact.
• Prompt-injection hardening — all user text is delimiter-wrapped as untrusted data before reaching an LLM; measured against adversarial cases in the eval suite.
• Output sanitization — user-derived text is entity-escaped and length-capped before rendering to Block Kit (no <!channel>/mention/link injection).
• Secrets hygiene — no credentials in the repo; local runtime uses .env (gitignored), the hosted service uses Render environment variables, and CI uses encrypted GitHub Actions secrets.