Hearth: a self-hosted household AI assistant tested for privacy and prompt-injection defence
The Hearth project is a private assistant for a household that answers questions from manuals and papers, citing sources and pages, and runs entirely on one home computer using Ollama, nomic-embed-text, faster-whisper and Django. In tests across eleven sample documents and 28 questions, models from 4 to 27 billion parameters answered accurately and reliably reported 'not found' when a document lacked the answer. Memory size determines which model can run, with the 9-billion-parameter model as the default. The team found that prompt wording could not stop small models from obeying hidden instructions planted in documents, and that placing such text in the system message made it more likely to be followed. Because small models remained vulnerable, Hearth removes sentences addressed to AI assistants in code when a document is added, which blocked the planted instructions in every test.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in