If you discover a security vulnerability in this project, please report it responsibly.
Do NOT open a public GitHub issue for security vulnerabilities.
Instead, email: shane@aianalystlab.ai
Include:
- Description of the vulnerability
- Steps to reproduce
- Potential impact
- Suggested fix (if you have one)
You should receive a response within 48 hours. We will work with you to understand and address the issue before any public disclosure.
This policy covers:
- The AI Analyst repository code
- Configuration files and templates
- Setup scripts
- Data handling and connection logic
This policy does NOT cover:
- Claude Code itself (report to Anthropic)
- MotherDuck (report to MotherDuck)
- Third-party dependencies (report to their maintainers)
- Never commit
.claude/mcp.jsonwith real tokens (use.claude/mcp.json.example) - Never commit connection templates with credentials (use
.yaml.examplefiles) - Never share your MotherDuck token publicly
- Review
.gitignorebefore pushing to ensure no sensitive data is tracked
Hooks. .claude/settings.json configures two Claude Code PostToolUse hooks that run after
tool calls in this project:
.claude/hooks/log-action.shappends one JSON line per tool action toworking/action_log_<date>.jsonl. This is the provenance trail thetraceskill reads, so "where did this number come from" points at logged actions, not a reconstructed story..claude/hooks/log-snowflake-query.shrecords queries made through the Snowflake MCP tool into the query log (requiresjq; exits silently without it).
Both write local files only and never block Claude. To turn them off, delete the hooks block
in .claude/settings.json.
Data handling. Your data is read and processed locally (pandas, DuckDB, or your warehouse connection). What leaves the machine:
- Your prompts and the context Claude reads (which can include query results) go to the Claude API as part of using Claude Code.
- Exports you explicitly invoke: Google Docs / Slides / Drive, Notion, Slack, email.
Nothing else. In particular, chart images for Docs and Slides are uploaded to your own Google
Drive through the chart-to-drive skill; no public file hosts are used anywhere in this repo, and
scripts/repo_lint.py fails CI if one is introduced.
Credentials. Warehouse credentials live in .env (gitignored) or your OS keychain, never in
tracked files; connection_templates/*.yaml.example and .env.example show the shape. The
analyst never prints credentials to the terminal, files, or logs.