Google DeepMind Strengthens Team to Secure AGI

Google DeepMind Strengthens Its Security Team for AGI

Google DeepMind, Google's artificial intelligence research branch, is expanding its team dedicated to the security and alignment of artificial general intelligence (AGI). This initiative is led by Rohin Shah, a recognized expert in the field, who aims to minimize the existential risks posed by advanced AI systems. Shah shares his insights and thoughts through a popular podcast hosted on the 80,000 Hours platform, where he discusses critical issues related to AI.

The AGI security and alignment team at Google DeepMind is not looking for a specific type of candidate, but rather exceptional individuals who can add significant value to the team. The open positions are varied, ranging from research engineering to pure science, and the team is flexible regarding the workplace. However, it is noted that most current team members find they are more productive when based in San Francisco or London. Interested candidates can apply for positions in the United States or the United Kingdom after consulting the tips provided to optimize their applications.

Many basic facts about why the AGI security and alignment team is a good place to work and about their research approach have not changed much since a post in 2025. The team particularly focuses on the risks of more severe harm from more advanced AI than the rest of Google DeepMind. Their work includes AGI alignment, defense against misaligned deployments, and support for coordinated security. A recent summary of their work highlights their hands-on approach, particularly in the areas of deep alignment and robustness testing, which are still evolving. These areas are relatively new for the team and have been developed in response to the increasing capabilities of AI systems.

Five-minute presentations during their AGI security course provide insight into their overall alignment strategy and research directions. These presentations are designed to give participants a clear understanding of the challenges and potential solutions in the field of AGI security.

A Clear Distinction from Other Security Teams

Google DeepMind's technical security team stands out due to its unique approach compared to other companies like Anthropic and OpenAI. It focuses on several key aspects that differentiate it:

• Technical Facilitators for Governance: AI governance requires tailored, standardized, and robust technical solutions to address potential abuses. Google DeepMind anticipates these needs to better support governance. This involves developing tools and methods that can be applied consistently across different AI companies.

• Preference for Knowledge over Engagement: Rather than engaging in sterile debates, the team seeks to produce tangible evidence to resolve factual disagreements and develop execution techniques. This means they emphasize the pursuit of truth and understanding the facts to guide their actions.

• Focus on Future Systems that are Much More Capable: The team is preparing for much more capable AI systems, based on conceptual arguments regarding future capabilities and risks. They believe that much of the current work on AI systems will not be applicable to systems that will pose significant risks in the future.

• Consistent Holistic Approach: Google DeepMind aims to integrate its work into a holistic approach, illustrated by detailed technical roadmaps like "An Approach to Technical AGI Safety and Security." These documents serve as guides to direct their research and development efforts.

To be clear, Google DeepMind does not claim that these aspects are unaltered goods. Engagement is obviously important in certain circumstances; it would be detrimental for everyone to focus solely on knowledge. Their focus on future systems has likely led them to miss out on impactful opportunities with current AI systems. Their main reason for writing this is to select candidates who are broadly enthusiastic about this approach and who will likely fit better into the AGI security and alignment team.

Reasons to Join Google DeepMind

Working within the AGI security and alignment team at Google DeepMind offers the opportunity to contribute to major impactful work. The team receives significant support from management, with access to cutting-edge models and substantial computing resources, enabling them to conduct security research that is difficult to achieve elsewhere. This advanced research infrastructure allows the team to push the boundaries of what is possible in AI security.

The team is composed of talented members, and those who appreciate external research are encouraged to undertake consulting projects. This openness helps to strengthen the capabilities of the AGI security field, with many alumni now working at companies like Anthropic, OpenAI, and UK AISI. This diversity of experience enriches the team and contributes to a more holistic approach to AGI security.

Team culture is another major asset, fostering collaboration and listening to objections, which allows for effective coordination around a shared mission. Team members are encouraged to express their ideas and actively contribute to discussions, creating a dynamic and innovative work environment.

Team members who appreciate external research are encouraged to engage in external consulting. Google DeepMind often has research project ideas that are not particularly advantaged by being inside a cutting-edge lab, so it may simply be preferable to carry them out externally. They also view this as a substantial investment in the capabilities of the AGI security field, and they are pleased to see that many alumni work full-time in security, some at Google DeepMind, but many also at Anthropic, OpenAI, UK AISI, and more.

Finally, although everyone says it, the team culture is genuinely excellent. Team members understand the goal of the project they are working on; any team member can raise an objection and will be heard. People are incredibly helpful and generous with their time. Because they genuinely share a mission, they can coordinate in ways that other teams struggle to achieve.

Profile of Desired Candidates

Google DeepMind seeks to recruit exceptional individuals who can adapt to various roles to maximize their impact. Candidates must share a common vision regarding the importance of AGI and the risks it poses while being self-sufficient and capable of working in a team. Alignment with the mission is crucial, as it allows the team to work coherently and effectively towards common goals.

While research experience is not mandatory, engineering skills are required. The emphasis is on productivity with high-performing coding agents, and engineering interviews allow the use of these tools. This reflects the importance of technological innovation in the team's work.

Ultimately, Google DeepMind wants to recruit excellent individuals and shape the team to enable them to have the most impact. That said, they particularly seek:

• Mission Alignment: They accomplish a lot because they are on the same wavelength. Members of the AGI security and alignment team adhere to the idea that AGI could arrive soon, that it creates unprecedented and potentially existential risks for humanity, and that they have a responsibility to act. Of course, they are not unanimous: there are significant disagreements within the team regarding the details and the best plan.

• Autonomy: They value proactive individuals. They work in a field with great uncertainty and difficult trade-offs. People who can be comfortable with uncertainty and make the best choices with the cards they have are very valuable to them. They need to accomplish tasks and execute quickly.

• Teamwork: Their team works well together because they appreciate and respect each other, and because they put the needs of the team first. They seek individuals who will help them continue in this direction.

You do not necessarily need to be an experienced researcher; they have certain roles that rely more on engineering skills. However, as a researcher, you should be a reasonably competent engineer. Note, however, that they primarily care about your productivity with high-performing coding agents. All their engineering interviews allow the use of coding agents (and they expect that most people who choose not to use coding agents will fail).

Recruitment Process

Applications can be submitted for positions in the United States or the United Kingdom, accompanied by a note explaining why the candidate would be a good fit and why they wish to join the team. After a preliminary screening of resumes, an initial interview is conducted, followed by remote interviews for a select number of candidates. Finally, shortlisted candidates undergo matching interviews with the team. The process is designed to identify candidates who not only possess the necessary technical skills but also share the team's values and vision.

Candidates can submit their applications here (for the United States) or here (for the United Kingdom). They strongly encourage including a note explaining in one paragraph why you would be a good fit and a paragraph on why you want to work with them. Use simple and direct language. Bullet points are acceptable and even recommended. Real reasons yield much better results than corporate jargon. This will be read by a real person, and it’s someone who will be disappointed by jargon and empty phrases. You can use the "cover letter" field of the form to upload this.

They then proceed to a selection of resumes and candidate notes. It is very helpful for you to highlight in your resume elements that directly speak to the things they are looking for. It can be particularly challenging to assess mission alignment solely from resumes, so if you do not have professional experience that clearly addresses the extreme risks of ASI/AGI, it is very important to explain why you wish to pivot towards this.

After the resume selection, there is a preliminary interview where they ask you about a few things that matter to them. This interview is short and not designed to reward excessive practice. There are no leetcode-style questions.

Then, a relatively small number of candidates undergo a "remote interview" with about four interviews on various topics. All engineering interviews allow unlimited assistance from agents.

Finally, a very small number of candidates are invited to a series of matching interviews with the team.

They work hard to reduce the total time of this process, but it typically takes at least a few months from start to finish.

Future Projects of Google DeepMind

The team plans to devote an increasing share of its efforts to putting its research into practice, particularly through deep alignment and control. Control measures, in particular, are expected to become essential for security in the near future. So far, research has focused on the monitorability and transparency of CoT, but the team plans to broaden its scope to seize new opportunities. This expansion of research efforts reflects Google DeepMind's commitment to staying at the forefront of AGI security and anticipating future challenges.

You can read their general approach to AGI safety and security. They wrote it over a year ago, but it has not changed much in its broad outlines. They anticipate that an increasing fraction of their efforts will be devoted to putting their research into practice, primarily through deep alignment and control. They expect that their control measures, in particular, will become essential for security in the near future.

So far, most of their research on control has focused on the monitorability of CoT and the preservation of CoT transparency. They believe they have exploited many of the easy opportunities in this area. In the future, they plan to increase their focus on the full range of control measures to ensure the safety of advanced AI systems.