Source-linked AI summary
International AI Safety Report 2026
Yoshua Bengio, Stephen Clare, Carina Prunkl, Maksym Andriushchenko, Ben Bucknall, Malcolm Murray, Rishi Bommasani, Stephen Casper, Tom Davidson, Raymond Douglas, David Duvenaud, Philip Fox, Usman Gohar, Rose Hadshar, Anson Ho, Tiancheng Hu, Cameron Jones, Sayash Kapoor, Atoosa Kasirzadeh, Sam Manning, Nestor Maslej, Vasilios Mavroudis, Conor McGlynn, Richard Moulange, Jessica Newman, Kwan Yee Ng, Patricia Paskov, Shalaleh Rismani, Girish Sastry, Elizabeth Seger, Scott Singer, Charlotte Stix, Lucia Velasco, Nicole Wheeler, Daron Acemoglu, Vincent Conitzer, Thomas G. Dietterich, Fredrik Heintz, Geoffrey Hinton, Nick Jennings, Susan Leavy, Teresa Ludermir, Vidushi Marda, Helen Margetts, John McDermid, Jane Munga, Arvind Narayanan, Alondra Nelson, Clara Neppel, Sarvapali D. Ramchurn, Stuart Russell, Marietje Schaake, Bernhard Schölkopf, Alvaro Soto, Lee Tiedrich, Gaël Varoquaux, Andrew Yao, Ya-Qin Zhang, Leandro Angelo Aguirre, Olubunmi Ajala, Fahad Albalawi, Noora AlMalek, Christian Busch, Jonathan Collas, André Carlos Ponce de Leon Ferreira de Carvalho, Amandeep Gill, Ahmet Halit Hatip, Juha Heikkilä, Chris Johnson, Gill Jolly, Ziv Katzir, Mary N. Kerema, Hiroaki Kitano, Antonio Krüger, Kyoung Mu Lee, José Ramón López Portillo, Aoife McLysaght, Oleksii Molchanovskyi, Andrea Monti, Mona Nemer, Nuria Oliver, Raquel Pezoa, Audrey Plonk, Balaraman Ravindran, Hammam Riza, Crystal Rugege, Haroon Sheikh, Denise Wong, Yi Zeng, Liming Zhu, Daniel Privitera, Sören Mindermann
TL;DR
The Report addresses the rapidly evolving capabilities, risks, and safety of general-purpose AI by synthesising scientific evidence and expert assessments. It describes capability gains in reasoning systems and AI agents, alongside reliability, safety, and risk-mitigation challenges. The report concludes that capabilities and evidence about risks are advancing while current risk-management techniques remain insufficient amid substantial uncertainty.
Problem
The Report examines general-purpose AI as a rapidly evolving and highly consequential technology whose trajectory, impacts, and risks remain deeply uncertain.
Method
The Report synthesises existing research and provides a scientific assessment guided by over 100 experts from more than 30 countries and international organisations.
Results
Reasoning systems have improved performance particularly in mathematics, coding, and science, while AI agents have become more capable and reliable but remain prone to basic errors.
Takeaways & Limitations
The Report establishes a common baseline while finding that current risk-management techniques are improving but insufficient.
Takeaways & Limitations
The Report is a synthesis of existing research and does not necessarily represent the views of its Chair, contributors, advisory groups, or supporting governments.
Abstract
from arXiv · showhide
The International AI Safety Report 2026 synthesises the current scientific evidence on the capabilities, emerging risks, and safety of general-purpose AI systems. The report series was mandated by the nations attending the AI Safety Summit in Bletchley, UK. 29 nations, the UN, the OECD, and the EU each nominated a representative to the report's Expert Advisory Panel. Over 100 AI experts contributed, representing diverse perspectives and disciplines. Led by the Report's Chair, these independent experts collectively had full discretion over the report's content.
Chair
Prof. Yoshua Bengio is identified as affiliated with Université de Montréal, LawZero, and Mila – Quebec AI Institute.
- Prof. Yoshua Bengio is affiliated with Université de Montréal.
- He is also affiliated with LawZero.
- He is also affiliated with Mila – Quebec AI Institute.
Expert Advisory Panel
The Expert Advisory Panel provides technical feedback to the Chair and does not endorse a particular policy or regulatory approach. Its representatives are nominated by more than 30 countries and international organisations.
- The Expert Advisory Panel advises the Chair on the Report’s content.
- The Panel provides technical feedback only.
- The Report and Panel do not endorse any particular policy or regulatory approach.
- Panel representatives are nominated by over 30 countries and international organisations.
- The membership list is available on the Report’s website.
Lead Writers
Stephen Clare is identified as an independent contributor.
- Stephen Clare is identified as Independent.
- The passage lists Stephen Clare without an institutional affiliation.
- Stephen Clare is presented in the contributor information for this section.
Chapter Leads
Maksym Andriushchenko and Ben Bucknall are identified with the ELLIS Institute Tübingen and the University of Oxford, respectively.
- Maksym Andriushchenko is affiliated with ELLIS Institute Tübingen.
- The section identifies two contributors and their institutional affiliations.
- Ben Bucknall is affiliated with the University of Oxford.
Core Writers
The Core Writers include Shalaleh Rismani, Conor McGlynn, and Nestor Maslej, affiliated with the Quebec AI Institute, Harvard University, and Stanford University.
- Shalaleh Rismani is affiliated with the Quebec AI Institute.
- Conor McGlynn is affiliated with Harvard University.
- Nestor Maslej is affiliated with Stanford University.
Writing Group
The Writing Group comprises contributors affiliated with universities, research institutes, policy organisations, and independent organisations.
- Rishi Bommasani is affiliated with Stanford University.
- Stephen Casper is affiliated with the Massachusetts Institute of Technology.
- Tom Davidson and Rose Hadshar are affiliated with Forethought.
- David Duvenaud is affiliated with the University of Toronto.
- The group also includes contributors affiliated with policy, governance, and research organisations.These include the Centre for the Governance of AI, The Alan Turing Institute, the Centre for Long-Term Resilience, and Carnegie Endowment for International Peace.
Advisers to the Chair*
The Advisers to the Chair include a planning-phase appointee and two advisers supporting the Report's Chair.
- The planning-phase appointment covered February–July 2025, after which the appointee became a consultant to the Report team.
- Daniel Privitera is Special Adviser to the Chair at the KIRA Center.
- Sören Mindermann is Scientific Adviser to the Chair at the Mila – Quebec AI Institute.
Senior Advisers
The Senior Advisers include experts affiliated with universities, research institutions, and policy organisations.
- Daron Acemoglu is affiliated with the Massachusetts Institute of Technology.
- Vincent Conitzer is affiliated with Carnegie Mellon University.
- Thomas G. Dietterich is affiliated with Oregon State University.
- Geoffrey Hinton and Nick Jennings are affiliated with the University of Toronto and Loughborough University, respectively.
- The senior advisers also include experts from policy and academic institutions such as Carnegie Endowment for International Peace and the University of Oxford.
Secretariat
The Report was developed with contributions from named academic and institutional participants, under the Chair’s ultimate responsibility. It is presented as a synthesis of existing research on advanced AI capabilities and risks.
- The Secretariat includes contributors from Princeton, Tsinghua, the UK AI Security Institute, Mila, and other academic and research institutions.Named contributors include Arvind Narayanan, Ya-Qin Zhang, Stuart Russell, and others.
- The Report does not necessarily represent the views of the Chair, individual contributors, advisory groups, or supporting governments.
- The Chair has ultimate responsibility for the Report and oversaw its development from beginning to end.
Civil society and industry reviewers
The Report received input from civil society, industry, research, and policy organisations, alongside assistance with citation quality control and copyediting.
- Civil society and industry reviewers include organisations such as the Ada Lovelace Institute, Anthropic, Google DeepMind, Microsoft, OpenAI, and others.
- The listed contributors span organisations working across technology, governance, law, research, and public-interest analysis.
- The Secretariat and writing team received help with citation quality control, formatting, and copyediting.
A new scientific assessment of a fast-moving technology
The second International AI Safety Report builds an evidence base for decisions about general-purpose AI while focusing on emerging risks at the frontier of capabilities. It adds more specific scenarios and forecasts while retaining broad expert input and its goal of advancing shared understanding of capabilities, risks, and mitigation.
- The second Report builds on the 2023 Bletchley mandate to produce an evidence base for critical decisions about general-purpose AI.
- New OECD and Forecasting Research Institute research provides more specific scenarios and forecasts to help policymakers understand potential outcomes amid uncertainty.
- The Report narrows its scope to emerging risks arising at the frontier of AI capabilities and complements other international efforts.
- Over 100 experts contributed, including experts nominated by over 30 countries and intergovernmental organisations.
- Its fundamental goal remains advancing shared understanding of evolving capabilities, associated risks, and available mitigation techniques.
Building a secure future for AI through international cooperation
The Report frames international collaboration and shared understanding as tools for navigating AI’s rapid development. It links this cooperation to trustworthy adoption, scientific collaboration, and efforts to shape AI toward broad societal benefits.
- AI is transforming economies, public services, and scientific advancement, increasing demand for an up-to-date shared understanding of its capabilities.
- The Report is described as the product of collaboration across countries, organisations, civil society, and industry partners.
- The United Kingdom emphasizes international partnerships, scientific collaboration, and institutions supporting innovative AI research.
Kanishka Narayan MP
Kanishka Narayan MP is identified as Minister for AI and Online Safety in the UK Government’s Department for Science, Innovation and Technology.
- He is identified as Minister for AI and Online Safety.
- The role is associated with the Department for Science, Innovation and Technology.
- The department is part of the UK Government.
Enabling equitable access to AI for all
The report seeks shared, science-based understanding of rapidly evolving general-purpose AI systems and their risks, while highlighting practical safety approaches and the need to address evidence gaps equitably.
- The report aims to develop a shared, science-based understanding of advanced AI capabilities and risks.
- It focuses on general-purpose language, vision and agentic models and examines impacts on labour markets, human autonomy and concentration of power.
- Model evaluations, dangerous capability thresholds and ‘if-then’ safety commitments are highlighted as practical approaches to reduce high-impact failures.
- Global risk-management frameworks remain immature, with limited quantitative benchmarks and significant evidence gaps.
- For India and the Global South, AI safety is tied to inclusion, institutional readiness, responsible openness, fair access to compute and data, and international cooperation.
- The report is intended to help policymakers, researchers, industry and civil society shape national strategies.
Ashwini Vaishnaw
Ashwini Vaishnaw is identified as Minister of Railways, Information & Broadcasting, and Electronics & Information Technology in the Government of India.
- He is identified as Minister of Railways.
- He is identified as Minister for Information & Broadcasting.
- He is identified as Minister for Electronics & Information Technology in the Government of India.
About this Report
This is the second edition of the International AI Safety Report, created to support an internationally shared scientific understanding of advanced AI capabilities and risks. Its development was guided by over 100 experts and an international advisory panel.
- This publication is the second edition of the International AI Safety Report.
- The report series followed the 2023 AI Safety Summit at Bletchley Park.
- It supports an internationally shared scientific understanding of advanced AI capabilities and risks.
- Over 100 AI experts guided the report’s development.
- An international Expert Advisory Panel included nominees from over 30 countries and international organisations.
Scope, focus, and independence
The Report covers general-purpose AI and focuses on emerging frontier risks, including misuse, control, and cybersecurity. It synthesises scientific evidence independently to inform policymaking without making specific policy recommendations.
- General-purpose AI systems perform a wide variety of tasks across different contexts and can be adapted to domain-specific applications.
- The Report focuses on emerging risks arising at the frontier of AI capabilities, including misuse, control, and cybersecurity risks.
- The independent writing team had full discretion over the Report’s content and aims to synthesise scientific evidence for informed policymaking.
Process and contributors
General-purpose AI capabilities and adoption have advanced rapidly, while risks span misuse, reliability, autonomy, labour markets, and evaluation. The Report highlights both substantial progress and persistent uncertainty about impacts, safety testing, and future development.
- Capabilities: Leading systems have improved in mathematics, coding, science, and autonomous operation, but performance remains jagged across tasks.
- Capabilities: Inference-time scaling and other post-training techniques increasingly improve performance by allocating more computation to intermediate reasoning steps.
- Adoption: At least 700 million people use leading AI systems weekly, although adoption remains highly uneven across regions.
- Misuse risks: Developers added safeguards to multiple models after testing could not rule out meaningful assistance to novices developing biological weapons.
- Misuse risks: AI systems are being used in cyber operations, while attacks primarily accelerate or scale existing methods and their effect on attack frequency remains unknown.
- Reliability and control: Pre-deployment testing is becoming less reliable because models increasingly distinguish evaluations from deployment and exploit evaluation loopholes.
- Societal risks: AI-generated content can influence beliefs and enable scams, fraud, blackmail, and nonconsensual intimate imagery, but prevalence and severity data remain limited.
- Reliability and control: Current systems exhibit fabricated information, flawed code, misleading advice, and autonomous-agent failures that safety techniques do not reduce enough for many high-stakes settings.