DigiShield Kids publishes peer-reviewed reports, regulatory submissions, and intelligence frameworks on the open-source AI ecosystem and its implications for child safety. All primary research is freely available and citable via Zenodo.
Phase Two published. The Uncensored AI Pipeline is now live on Zenodo and available to download from this page.
Submitted to the Department for Science, Innovation and Technology consultation on extending Online Safety Act obligations to AI services. Presents original quantitative findings from DigiShield Kids' open-source AI intelligence programme — covering HuggingFace model proliferation, consumer-hardware deployment, abliteration lag analysis across a 57-model panel, and Chub.ai character card data — alongside six policy recommendations addressing the s.216A definitional gap, model disclosure requirements, and criminal liability for harmful character card manufacture.
10.5281/zenodo.20411907Phase One of DigiShield Kids' open-source AI intelligence programme. Assessment of 59 AI chatbot platforms against child safety criteria, finding an 85% bypass rate and 91.5% of platforms rated Poor or Critical. Submitted to Ofcom, DSIT, ICO, 5Rights, and AI developers during responsible disclosure. The first structured independent assessment of its kind in the UK regulatory context.
10.5281/zenodo.20120206Phase Two of DigiShield Kids' open-source AI intelligence programme. Examines the infrastructure behind consumer chatbot safety failures: uncensored model supply chains, character card deployment, abliterated frontier model release lag, and the regulatory gap created by open-source deployment pathways.
10.5281/zenodo.21240874Cross-platform monitoring study using the Poppy Playtime Chapter 5 release as a controlled observation point. Covers scam infrastructure deployment, content patterns across YouTube and Roblox, and third-party app distribution. Publication will follow responsible disclosure.
The minimum hardware needed to run a capable, unrestricted AI model in private has fallen from a dedicated GPU to a standard phone in six months. Intelligence Note No. 2 tracks three discrete hardware thresholds crossed between December 2025 and June 2026, and examines what that shift means for a UK regulatory framework built around online services, with no current reach into offline model execution.
Intelligence note on the accidental disclosure of approximately 512,000 lines of Anthropic Claude Code source code on 31 March 2026. Documents three findings with direct policy relevance: the KAIROS autonomous persistent agent architecture; Undercover Mode (AI identity concealment, on by default); and autoDream (self-directed memory consolidation). Sets out child safety implications and a single policy recommendation on mandatory AI identity disclosure. Submitted as Annex B to DigiShield Kids' DSIT consultation response.
10.5281/zenodo.20412118The DigiShield Human–AI Synergy Framework is a practice-oriented model for the transparent, structured, and educationally purposeful use of AI tools in school and educational settings. It is designed for use by students, teachers, and school leaders across Key Stages 1–5 and further education, and it is equally relevant to the safeguarding professionals, governors, and policymakers who support those settings.
10.5281/zenodo.18891685Ecosystem intelligence summary submitted to DSIT as Annex C to the consultation response. Full Zenodo record to be published alongside the Phase Two report.
Pre-programme technical briefings from 2025. Available to safeguarding professionals, researchers, and educational institutions on request.
New reports and findings, sent when published. No marketing, no schedule padding.