• Your ANTHROPIC_API_KEY (read from env or constructor arg)
• The decision text you pass in
• The voice verdicts Claude returns

• Persist any data anywhere by default (Brier audit module is purely in-memory until you persist)
• Send any telemetry or analytics
• Log anything outside what you explicitly print

If you find a security issue (e.g. accidental key logging, prompt injection that exfiltrates context, dependency CVE), please do not open a public GitHub issue.

Email: xji1@mail.yu.edu

Include:

• A minimal reproduction
• Your assessment of severity
• Any disclosure deadline you'd prefer

I'll respond within 72 hours.

@anthropic-ai/sdk is the only runtime dependency. Audit it however you'd audit any other LLM client.

• Never commit a real ANTHROPIC_API_KEY to a repo
• Use .env / .env.local + add to .gitignore
• Rotate keys if they touch chat / screenshots / pastebin
• For server-side use in vibex-style deployments, store in Vercel env vars (uncheck "Sensitive" flag — see vibex AGENTS.md for the known Vercel Sensitive gotcha)

The Claude call accepts user-provided decision and context strings. These flow into the model. Treat any output as potentially adversarial — don't render verdicts as HTML, don't eval() them, and don't use them to drive downstream automated actions without HITL review.