The design philosophy behind Vulcan Veil

Even as a pro user of different generative AI tools for writing computer code and doing research, I find it is still very difficult to orchestrate these tools in a safe manner. Hence the need for Vulcan Veil, a desktop and web chat harness intended to help with the real tasks analysts need to do, while keeping safety a priority.

So while building it, I have these goals in mind:

• Human-in-the-loop: I expect a human not only to drive the work being done, but also to be a critical reviewer of that work. Hence, by default, any work done needs to be reviewed.
• Batteries included: It is difficult for many non-power users to set up their computer to use generative AI tools efficiently, which in turn becomes an actual vector for malware. I have built Vulcan Veil with many sandboxed tools included by default in the desktop application (an isolated web browser, a file editor for PDF, Word, and Excel, sandboxed Python with many of the scientific libraries, and conversion of documents into a store that AI agents can easily traverse).
• Privacy first: I use only true zero-data-retention model servers hosted in the US (Baseten, Databricks, and AWS GovCloud). Even the current frontier model providers (OpenAI and Anthropic) save chats in ways that are not reasonable for individuals who are focused on privacy.
• Safety first: I have built the tool with safety in mind first: tools to redact sensitive information locally, to prevent environment secrets from being sent to AI tools and saved in code, and an isolated environment to run code separate from the rest of the computer.
• Whitelist first: The concept behind whitelisting is that you only allow the machine to access the specific resources you give it access to. Almost all of the agentic AI tools now do not limit access by default (they can access all of your files, your computer, and any command you can personally run). Whitelisting in Vulcan Veil means only specific files and folders can be read and edited, and specific tools can be used only if you turn them on.

This will cause some friction for power users, but these limits can ultimately be turned off if you choose to do so (for personal use; for team accounts, it is controlled at the account holder level).

While working with my consulting firm, Crime De-Coder, as well as writing technical books on programming and AI, I have come to the realization that many individuals need much more help getting started using AI tools in a safe way. It is very easy to accidentally send all of your data to a provider, or to expose your internal code to an exfiltration attack.

This tool is intended to be an alternative to desktop tools like Microsoft Copilot, Claude Desktop, or Codex, in which controlling risky behavior tends to be very difficult.