SShortSingh.
Back to feed

Hearth: a self-hosted household AI assistant tested for privacy and prompt-injection defence

0
·4 views

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.

Read the full story at DEV Community

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

Related stories

0
ProgrammingDEV Community ·

Designing the Order Flow for a Custom Children's Book

The article argues that personalised products depend heavily on the order flow, since buyers pay before seeing the result. It uses The Personalized Pet Story Book for Kids, which features a child and their pet and comes in Mini (₹797) and Regular (₹1,444) sizes, as a case study. The author recommends clear photo guidance at upload, separate fields for child and pet names, an explicit spelling confirmation step, and an upfront delivery timeline of about 7–10 business days after preview approval. The piece stresses that the WhatsApp preview builds buyer trust, and suggests tracking drop-off between upload and checkout as a key metric.

0
ProgrammingDEV Community ·

Keeping related-link blocks stable across publishing platforms

A DEV Community post discusses how to keep related-link blocks consistent when content is published on multiple platforms. The recommended approach is to generate the link block once and append it after the article body. This matters because platforms vary in how they support markdown and HTML. The post links to its source repository, a Mbin instance, and the author's DEV profile. The article text provided is brief, so further details of the proposed technique are not available.

0
ProgrammingDEV Community ·

Why QA Teams Reach Different Verdicts on the Same AI Output

Testers evaluating the same AI response can reach opposite conclusions, such as calling a food-photo answer of 'curry' a pass or a fail. The author argues the root cause is undefined grading rules rather than the model itself, including unclear units of correctness, no agreed status for partial answers, and ground truth that depends on individual labelers. Using a food recognition feature as an example, a disagreement over whether two identically named sandwiches count as a duplicate came down to comparing names versus comparing content. The recommended remedy is to write rules and borderline examples before testing, have teams grade a small sample independently, and settle disputes by considering the user's actual experience rather than seniority. Teams that cannot answer key grading questions in a sentence, the author warns, are likely to see their pass rates diverge.

0
ProgrammingDEV Community ·

Changing DNS providers: a 12-point checklist to keep your site and email running

Switching DNS providers is a migration rather than a simple settings change, because updated nameservers propagate gradually across resolvers worldwide over hours to days. Mismatches between old and new zones can cause intermittent failures for visitors and mail servers that are hard to diagnose. The guide recommends exporting the full existing zone, including easily overlooked records such as DMARC, DKIM, ACME challenge and SRV entries, and recording the current state as a baseline. It advises lowering TTLs ahead of the move, creating every record on the new provider, and diffing its authoritative answers against the baseline before switching. It also warns that DNSSEC requires a specific order of steps, since a mismatched DS record can make a domain fail validation, and suggests external uptime monitoring during the cutover.