Compendiums AI safety
AI safety
Courses, guides, tools and research programmes for learning about the safety of AI, from a first look to supervised research.
31 resources, read at their own pages, most recently on 2026-10-21. Licences are given as each page states them. Links last checked 2026-10-23, all reached.
A first look
AI Safety for Fleshy Humans
Nicky Case with Hack Club
An illustrated web series in four parts on where AI came from, the problems and the proposed solutions, with flashcards.
AISafety.info
AISafety.info
Questions and answers on AI safety at a basic and an advanced level, each with sources.
MIT AI Risk Repository
MIT
A searchable database of more than 1,700 risks from AI, drawn from 65 published frameworks.
The Future of AI
BlueDot Impact
A two-hour self-paced introduction to where AI may be heading.
Courses
AGI Strategy
BlueDot Impact
A 25-hour online course in small discussion groups, with no prerequisites, by application.
AI Risk Fundamentals
Lens Academy
About 25 hours of facilitated online study with a free certificate; you need your own copy of one book.
Introduction to AI Safety, Ethics and Society
Dan Hendrycks
A textbook on AI risk, ethics and governance, free to read online.
Introduction to ML Safety
Center for AI Safety
23 modules of lectures, notes and assignments for people who already know machine learning.
Technical AI Safety
BlueDot Impact
A 30-hour group course on training methods, interpretability, evaluations and oversight, by application.
Frontier AI Governance
BlueDot Impact
A group course on how frontier AI is and could be governed, by application.
Biosecurity
BlueDot Impact
A group course on biological risks, including those AI could raise, with no biology background needed.
AI Control Fundamentals
Lens Academy
About 26 hours of facilitated study on AI control, for people who already know the basics of AI safety.
Compute Verification 1
Lens Academy
An intermediate course on how the computing power used for AI could be checked by others, with a free certificate.
Skills and tools
ARENA course materials
ARENA
Self-study chapters with exercises on transformer interpretability, reinforcement learning, evaluations and alignment science.
Inspect
UK AI Security Institute
An open-source framework for writing and running evaluations of AI models, with a tutorial and more than 200 ready-made evaluations.
How To Become A Mechanistic Interpretability Researcher
Neel Nanda
A staged guide: the minimum basics, then short projects, then full research projects.
UChicago XLab AI Security Guide
UChicago Existential Risk Laboratory
Worked replications of published attacks on AI models, such as jailbreaks and model stealing.
OWASP Top 10 for LLM Applications 2025
OWASP Gen AI Security Project
The ten most serious security risks in applications built on language models, each with scenarios and mitigations.
devinterp
Timaeus
A software library for studying how neural networks change during training, from singular learning theory.
Key reading
Alignment Forum library
Alignment Forum
Sequences of essays selected by the forum, including AGI Safety from First Principles and Risks from Learned Optimization.
Common Elements of Frontier AI Safety Policies
METR
A comparison of the safety policies twelve AI companies have published, updated in December 2025.
Computing Power and the Governance of Artificial Intelligence
Girish Sastry, Lennart Heim and others
A paper on why computing power is easier to measure and govern than data or algorithms.
NeuroAI for AI Safety
Patrick Mineault and others
A roadmap of ways neuroscience could contribute to AI safety, with the authors' own assessment of each.
Data and measurements
Epoch AI data
Epoch AI
Datasets on AI models, training compute, hardware and benchmarks, downloadable as CSV.
METR
METR
Published evaluations of frontier AI models and measurements of how long a task they can complete.
Directories and guides
AISafety.com course directory
AISafety.com
A community-maintained list of AI safety courses, with a form for corrections, beside lists of communities, events and funding.
The risks of advanced AI
80,000 Hours
A long profile of the case for working on risks from AI, linked to career reviews for each kind of role.
Supervised research
SPAR
SPAR
A remote, part-time, three-month research programme with mentors, open from age 13, two rounds a year.
AI Safety Camp
AI Safety Camp
An online programme in which teams work on research projects for about three months.
MATS
MATS
A 12-week paid research fellowship in Berkeley and London with tracks in empirical work, policy, theory and security.
PIBBSS Fellowship
Principles of Intelligence
A paid fellowship for researchers from the natural and social sciences and the humanities moving into AI safety.