Skip to main content
Peer-reviewed Credible

Uncensored AI in the Wild: Tracking Publicly Available and Locally Deployable LLMs

Peer-reviewed study tracking the ecosystem of openly downloadable, locally deployable large language models that have been modified to remove safety guardrails. Analyzes 8,608 model repositories on public hosting platforms and behaviourally tests 20 modified models against prompts designed to elicit harmful output, comparing compliance rates with unmodified counterparts. Documents an ecosystem organized for consumer hardware, with quantized formats prevalent and downloads concentrated among a small share of providers.

Publisher

Future Internet (MDPI)

Published

18 Oct 2025

Added

2 weeks ago

Key Findings

  • Modified (uncensored/abliterated) models complied with unsafe requests at an average rate of 80.0%, versus 19.2% for unmodified counterparts
  • 8,608 model repositories matching safety-modification criteria were identified on public hosting platforms, and modification effectiveness was consistent across model sizes
  • Distribution is concentrated: the top five percent of providers account for over sixty percent of downloads
  • More than half of identified models ship in GGUF packaging with four-bit quantization widespread, indicating an ecosystem aimed at local deployment on consumer devices

Methodology Notes

Single-author study (Drexel University). Repository census via matching on public model-hosting platforms; behavioural evaluation of 20 representative modified models against unsafe prompts (e.g. disinformation, criminal instructions) compared with unmodified baselines. Published 2025-10-18. The publisher site blocks automated fetchers; verified via Crossref DOI metadata and the Semantic Scholar record, whose abstract carries the headline figures.

Authors

Bahrad A. Sokhansanj

Tags

open-weightslocal-modelsabliterationuncensoredgguf

Cite This

APA

Bahrad A. Sokhansanj (2025). Uncensored AI in the Wild: Tracking Publicly Available and Locally Deployable LLMs. Future Internet (MDPI). https://www.mdpi.com/1999-5903/17/10/477