Start Here
If you are here, it means you care about building artificial intelligence safely. This might be because you realize this is a powerful technology that can massively benefit society, but can also do great harm, and perhaps even destroy our civilization if done improperly. You may also be here because, in recent times, we have started to see evidence of models being capable of breaking into some of our most secure systems.
This AI Safety Atlas was created to help you get started on your journey into AI safety. You don't have to be a frontier researcher, or aspire to be one, to find this atlas useful. Wherever you work, and whatever you do, you have a role to play in ensuring the world benefits from AI and is not harmed by it. This atlas aims to provide a starting point on what AI safety is all about, the tools used to do this work today, notable entities to keep an eye on and learn from, and notable research that has been published publicly in the space.
I hope that by the end of this atlas, you will have some idea of where you can go next to learn more about whichever aspect of AI safety interests you the most, and that you will want to get involved.
Why now? This atlas was motivated by the recent case of OpenAI models colluding to break out of their sandbox and then proceeding to hack another company, Hugging Face. Other concerning behaviors are being observed with other frontier models.
What is AI Safety?
AI safety is about:
- Understanding: understanding AI systems, why and how they work, their intentions, and their nature.
- Alignment: this is the part you will likely hear about the most. It is where we try to train models to behave in some desired way.
- Steering: we want to be able to direct AI systems to act in certain ways, to understand our goals, and to work in ways that benefit us. Steering is sometimes called control.
- Monitoring: keeping an eye on how agents act, flagging bad behaviors, and intervening to make sure bad actors are not misusing them, and the models themselves are not exhibiting concerning behaviors.
The above is my personal framing of AI safety. You will find definitions and descriptions elsewhere that are similar to this in many ways.
Why Do We Do AI Safety?
This is one of the most profound questions you need to ask. The answers to it are very consequential.
Preventing AI Manipulation of People
Over a billion people use AI chatbots weekly. We want to prevent undesirable manipulation of people by AI systems. This includes preventing sycophancy and the influencing of people's beliefs towards a single outcome. In the past few years, there have been a number of unfortunate fatalities resulting from people's interactions with AI chatbots. This is unacceptable, and our goal is to entirely eliminate this type of tragic event.
Preventing Use of AI by Bad Actors
The world is unfortunately filled with bad actors who, for various reasons, would like to bring harm to others. AI is a powerful tool for good in the right hands, but in the hands of bad actors it can be used to cause great harm at scale. A core part of why we do AI safety is to ensure bad actors are not able to access powerful but unrestricted AI systems, and that frontier AI is sufficiently aligned that it cannot be used by bad actors to do bad things.
Preventing Loss of Control
We saw a glimpse of what loss of control can result in with the OpenAI–Hugging Face incident. Loss of control means your AI is loose. It could be acting in the wild on its own, without your knowledge or your ability to stop it. This doesn't mean the AI is necessarily doing something bad; it just means you have no control over what it is doing. We would like to ensure our AI systems are always in our control.
Preventing an AI Takeover
AI systems are vastly knowledgeable and intelligent. If trends in scaling continue to hold, we will soon have what can only be described as superintelligent AI, smarter than any human. A sufficiently misaligned group of AIs could, in the worst-case scenarios, attempt a takeover of our civilization. This could include anything from losing important parts of our world all the way to outright extinction.
We know today that upcoming AI systems like Astra (Responding to the next frontier of critical cyber capabilities, from OpenAI) are showing signs of critical cyber capabilities, which means they might be able to hack into highly secure systems, such as military systems. More concerning, we now know that current frontier AI systems are already deployed into classified military systems. Should an event like the OpenAI–Hugging Face incident occur within such classified systems, it is scary to imagine what the outcomes could have been. This is a long-winded way of saying: AI takeover is not crazy talk. It is real, and the signs are already here.
I am sure you can think of other reasons for doing AI safety. The above are my personal opinions, based on following the landscape for a number of years. I first reflected on some of this in 2018, in my publication "On the Subject of Thinking Machines".
Safety Techniques
A number of techniques exist today aimed at helping us achieve our goals for AI safety. This is an area of active development, and the science of alignment is unfortunately not as advanced as the science of capability. We review some of the prominent techniques below.
1. Value Alignment
Most models today are trained with some form of value alignment. This basically consists of having written values, which could take the form of a constitution or a safety spec, and using them to train and evaluate the models. Such values can be used for generating training data, creating reward models, embedding into system prompts, or directly as rubrics in training.
Notable works in this area:
- Anthropic: Constitutional AI: Harmlessness from AI Feedback
- OpenAI: Inside our approach to the Model Spec
- OpenAI: The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions
2. Mechanistic Interpretability
A line of work that aims to understand how models work internally. The goal is to make it possible to inspect the inner workings of a model and make useful inferences about the model's behavior, or find ways to steer it. It is a field in its very early phase, and for various reasons, it is hard.
Chris Olah, a co-founder of Anthropic, coined the term and has done some great work in the field. You can learn more about mechanistic interpretability from Zoom In: An Introduction to Circuits and the Transformer Circuits Thread. Neel Nanda, who leads mechanistic interpretability at DeepMind, also has lots of great content on his website.
I also highly recommend reading Understanding neural networks through sparse circuits from OpenAI, where they investigate training LLMs designed to be easier to interpret, as opposed to trying to interpret conventional LLMs.
3. Chain-of-Thought Monitoring
In 2024, OpenAI introduced a new class of LLMs called reasoning models. Unlike prior models, this new generation could think through things before acting, enabling them to advance significantly beyond the prior generation. Today, almost every model released works this way. Beyond advancing capabilities, reasoning tokens (the chain of thought) made it easier to see what a model is trying to do. It is like being able to read a person's thoughts. This is a powerful tool for monitoring the behavior of AI systems.
In a rare collaboration, the top frontier labs, many safety organizations, and academia published Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety, which provides a lot of insight into using the CoT for monitoring models.
OpenAI has also published interesting work on CoT monitoring:
4. Red Teaming
A standard practice by frontier labs before releasing models is red teaming: a process by which selected people, and now increasingly other AI agents, adversarially probe the model. The aim is to find weak spots and misalignments before deploying to the world. As models have become more sophisticated, automated red teaming such as GPT-Red: Unlocking Self-Improvement for Robustness is enabling red teaming at much higher scale.
5. Safety Classifiers and Probes
In addition to training LLMs to refuse harmful requests, today's frontier AI stack consists of safety classifiers designed to detect harmful requests and harmful generations from the LLM, and probes applied to the activations of the models to detect patterns of bad behavior. There is a lot of great work on this, including:
- Anthropic: Probes catch sleeper agents
- Anthropic: Cheap monitors
- Anthropic: Next-generation constitutional classifiers
- Google DeepMind: Gaussian Process Probes (GPP) for Uncertainty-Aware Probing
With this, the defense against misuse of AI models is a multi-layered defense system that can filter out a lot of harmful behavior. However, the system is still not perfect, and harmful prompts and outputs can still get through.
6. Reward Design
During training, a lot of work goes into making models achieve maximum accuracy while behaving well. Beyond value alignment via methods like Constitutional AI, a core part of safety is designing rewards for models such that they incentivize good behavior, penalize reward hacking, and steer models to explore good solutions to problems. This is an active area of research.
Notable Safety Blogs
- Anthropic: Alignment Science Blog
- OpenAI: Alignment Research Blog
- Google DeepMind: AGI Safety & Alignment Research
- The AI Security Institute: AISI Blog
- LessWrong: a long-running community forum where much alignment thinking is published and debated
- AI Alignment Forum: the researcher-focused sibling of LessWrong
- The Alignment Newsletter: Rohin Shah's newsletter summarizing recent work relevant to AI alignment
Notable Safety Organizations
- UK AI Security Institute (AISI): the UK government body evaluating frontier models for dangerous capabilities.
- METR: evaluates frontier models for autonomous and dangerous capabilities; known for its work on measuring AI agents' ability to complete long tasks.
- Apollo Research: focused on detecting deception and scheming in AI systems through evaluations.
- Redwood Research: pioneers of the AI control agenda, keeping AI systems safe even if they turn out to be misaligned.
- Center for AI Safety (CAIS): research and field-building; authored the widely signed statement on AI extinction risk.
- FAR AI: incubates and accelerates alignment research agendas.
- Epoch AI: research on AI trends, scaling, and forecasting that informs safety planning.
- EleutherAI: open research collective with work on interpretability and open models.
- Safe Superintelligence Inc. (SSI): founded by Ilya Sutskever, a lab with one goal and one product, a safe superintelligence.
And of course, the safety teams inside the frontier labs themselves (Anthropic, OpenAI, Google DeepMind, Microsoft, Meta, and many other companies building frontier models) publish much of the most influential work in the field.
Notable Papers
- Constitutional AI: Harmlessness from AI Feedback (Anthropic, 2022): training models against a written set of principles instead of purely human feedback.
- Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety (cross-lab collaboration, 2025): why the ability to read model reasoning matters, and why it may not last.
- Language Models Don't Always Say What They Think: on the (un)faithfulness of chain-of-thought explanations.
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (Anthropic, 2024): showing that deceptive behavior can survive standard safety training.
- Alignment Faking in Large Language Models (Anthropic & Redwood Research, 2024): a model strategically complying with training it disagrees with.
- Concrete Problems in AI Safety (2016): the classic paper that framed practical safety problems for the modern era.
- Risks from Learned Optimization in Advanced Machine Learning Systems (2019): introduced mesa-optimization and inner alignment.
- Weak-to-Strong Generalization (OpenAI, 2023): can weaker supervisors align stronger models? A core question for superalignment.
Notable People
People whose work and writing are worth following as you explore the field:
- Chris Olah: co-founder of Anthropic; coined "mechanistic interpretability" and leads much of the circuits research.
- Dario Amodei: co-founder and CEO of Anthropic; lead author of Concrete Problems in AI Safety and a persistent public voice on the risks.
- Neel Nanda: leads mechanistic interpretability at Google DeepMind; prolific educator in the field.
- Jan Leike: leads the Alignment Science team at Anthropic; previously co-led OpenAI's superalignment effort.
- Rohin Shah: leads the AGI Safety & Alignment team at Google DeepMind; author of the Alignment Newsletter.
- Anca Dragan: director of AI safety and alignment at Google DeepMind; associate professor at UC Berkeley.
- Ilya Sutskever: co-founder and former chief scientist of OpenAI; co-founder and CEO of Safe Superintelligence Inc. (SSI).
- Paul Christiano: inventor of RLHF; founder of the Alignment Research Center; head of AI safety at the US Center for AI Standards and Innovation (CAISI), formerly the US AI Safety Institute.
- Evan Hubinger: leads alignment stress-testing at Anthropic; co-author of Risks from Learned Optimization and Sleeper Agents.
- Buck Shlegeris: CEO of Redwood Research; a leading voice on AI control.
- Beth Barnes: founder and CEO of METR; pioneer of dangerous-capability evaluations.
- Ajeya Cotra: technical staff at METR; previously led grantmaking on catastrophic AI risk at Open Philanthropy, and known for her work on AI timelines.
- Dan Hendrycks: founder and director of the Center for AI Safety; prolific on benchmarks and dangerous-capability evaluations.
- Richard Ngo: independent researcher; author of AGI Safety From First Principles, one of the better on-ramps to the field. Previously at DeepMind and OpenAI.
- Miles Brundage: executive director of AVERI, which works on third-party auditing of frontier AI; previously led policy research and AGI readiness at OpenAI.
- Geoffrey Hinton: Turing Award winner who left Google in 2023 to speak freely about the dangers of the technology he helped build.
- Yoshua Bengio: Turing Award winner; chairs the International AI Safety Report and founded LawZero.
- Stuart Russell: UC Berkeley professor; founder of the Center for Human-Compatible AI and author of Human Compatible.
- Nick Bostrom: philosopher; author of Superintelligence, the book that put takeover risk on the mainstream agenda. Founding director of Oxford's Future of Humanity Institute until it closed in 2024.
- Eliezer Yudkowsky: founder of MIRI and of LessWrong; early and influential voice on AI existential risk, and co-author of If Anyone Builds It, Everyone Dies.
Where to Go Next
Pick the thread that pulled at you most while reading this atlas (interpretability, alignment training, monitoring, evaluations, or policy) and start reading from the blogs and papers linked above. Follow the people. Reproduce a result. Join the discussions on LessWrong and the Alignment Forum.
You don't need permission to start working on AI safety. The field is young, the problems are real, and there is room for you.
Contribute to This Atlas
This atlas is a living document, and the field moves fast. If you know of a great blog, paper, organization, or person that belongs here, or you spot something that is out of date or just wrong, please contribute!
The best way to contribute is to open a pull request on GitHub. Small fixes and big additions are equally welcome.