DontFeedTheAI Anonymizes Sensitive Data Before It Reaches Cloud LLMs

Summary

DontFeedTheAI is an open-source proxy for AI-assisted penetration testing and other workflows involving confidential data. It detects IP addresses, hostnames, organization names, credentials, hashes, tokens, API keys, and personally identifiable information, replacing them with surrogates before a request reaches a cloud LLM. Responses containing those surrogates are mapped back to the original values before being returned to the user. Detection combines a local Ollama model for entities such as hostnames, organization names, and credentials in prose with local regular-expression rules for structured secrets and identifiers, so sensitive data is intended to remain on the user’s machine. The proxy supports Claude Code, the OpenAI SDK, OpenRouter, Requesty, and other OpenAI-compatible clients, and can be deployed locally or through a VPS with an SSH tunnel. A wizard handles setup, engagement configuration, deployment, and client launch on Windows, macOS, and Linux. The repository includes an audit view that exposes original-to-surrogate mappings for debugging; the documentation warns that this tool should run only behind the SSH tunnel because it reveals the lookup table. It also provides 53 pipeline fixtures, integration tests requiring Ollama, and a regex-focused improvement loop for finding leaks and false positives. The project is released under the MIT license, while its threat-model documentation describes protections, limitations, and planned changes.