Fund Workbench is an early open-source project. Treat it as a review aid, not a secure document room.

• The local app has no database and no telemetry.
• Uploaded documents are processed in memory by the running Streamlit process.
• Deterministic mode sends nothing to a model provider.
• AI-assisted mode sends extracted source text to the configured OpenAI API account.
• AI requests set store=False; provider and organization retention settings still apply.
• Markdown and JSON exports are downloaded or written where the user chooses.

• Do not upload or transmit confidential material without authority.
• Never commit API keys or sensitive source documents.
• Use a secret manager or environment variable for credentials.
• Review every generated claim and memo before using it in a decision.
• Run the app locally when handling non-public data.

• PDF text extraction does not perform OCR.
• Prompt injection inside documents is treated as source text, but model behavior is not a security boundary.
• Source verification checks provenance within supplied documents; it does not establish that an independent source is accurate.
• The app does not provide authentication, tenant isolation, encryption at rest, or an audit log.
• Streamlit’s default deployment settings are not a substitute for an organization’s access controls.

Do not open a public issue for a vulnerability or include real sensitive data in a report. Contact the maintainer privately through the repository owner’s GitHub profile with a synthetic reproduction.