About
What this is
AI Safety Sandbox collects practical AI safety projects for people starting out. The projects come from published idea lists, credited on each page, and many are written up as short step-by-step guides. It was built for AI Safety Cardiff University (AISCU) (opens in a new tab).
How the matcher works
Each project is scored on five things:
- Theme (30%): whether it’s in a theme you picked. Projects that overlap with your theme count partly.
- How you like to work (20%): chatbots, code, writing, data, people or visuals.
- Experience (18%): guided starters for newcomers, larger projects if you’ve done some already.
- Background (17%): how well the skills it needs match your background.
- Time (15%): projects shorter than the time you have are fine; much longer ones score lower.
The highest-scoring projects are shown, with similar projects spread out so you don’t get near-duplicates. Each suggestion lists why it was picked.
The optional text box searches all projects using BM25 ranking. It corrects spelling mistakes (“intepretability”), completes partial words, and also looks for related ideas: a hand-written list of synonyms plus word links built from ConceptNet Numberbatch, so “flattery” finds sycophancy, “chips” finds compute governance and “bioweapons” finds biosecurity. If nothing fits well, the matcher says so and links to SPAR (opens in a new tab), AISafety.com (opens in a new tab) and free advisors.
Related links
Project pages link to related projects, further reading and similar mentored projects. These links are worked out when the site is built by comparing the text of every project and resource.
Privacy
The site uses no cookies, analytics or third-party services. Your matcher answers appear in the page address so you can share them. Completed steps are stored only in your browser.
Sources and credits
There are currently 108 projects, 228 mentored projects and 1,575 resources, drawn from:
- BlueDot Impact, Technical AI Safety Project (Unit 1: project ideas) (opens in a new tab): Five project ideas (Inspect Evals, emergent misalignment, side-effects of RLVR, beneficial capability evals, alignment faking) and core resources.
- BlueDot Impact, Biosecurity course (Unit 6) and the resources it links to (opens in a new tab): Five DNA-screening projects with their end-2025 status, a hundred biosecurity project ideas, dual-use guidance and rapid grants.
- AI Safety × Philosophy reading list: project ideas: Twenty-three projects that pair a small experiment with a philosophical argument (compiled for AISCU, November 2026).
- Technical AI Governance resource pack (opens in a new tab): Reading lists and research advice from the MIRI Technical Governance Team, and the MINT Lab AGI Governance Bibliography.
- SPAR (Supervised Program for Alignment Research) (opens in a new tab): 228 mentored research projects from the f26 round, retrieved 8 November 2026. Each links back to SPAR.
- AISafety.com Data API (CC BY 4.0) (opens in a new tab): Courses, programmes, funders, advisors, communities, organisations, media and volunteer projects, retrieved 8 November 2026.
- BlueDot blog (opens in a new tab): Posts on first open-source contributions, communications projects, a decision-making hackathon and rapid grants.
- Other sources: METR, Epoch AI, the Federation of American Scientists, the AI Incident Database, the MIT AI Risk Repository, OECD.AI, GOV.UK, the EU AI Act Explorer, Neuronpedia and the papers linked from each project.
Project write-ups paraphrase their sources, and every project page links to its original. Fonts are Inter and Space Grotesk (SIL Open Font License). Related-word search uses data from ConceptNet Numberbatch and wordfreq by Robyn Speer (CC BY-SA 4.0); the derived word list is shared under the same licence. The triangle pattern is inspired by Paul Ouwerkerk’s painting Calibration (The Oxford Variations, 2026).
Corrections
To suggest a project or report a broken link, contact AI Safety Cardiff University (AISCU) (opens in a new tab).