SafetySecurityTransparencycs.CY
International AI Safety Report 2025: First Key Update: Capabilities and Risk Implications
Yoshua Bengio, Stephen Clare, Carina Prunkl, Shalaleh Rismani, Maksym Andriushchenko, Ben Bucknall, Philip Fox, Tiancheng Hu, Cameron Jones, Sam Manning, Nestor Maslej, Vasilios Mavroudis, Conor McGlynn, Malcolm Murray, Charlotte Stix, Lucia Velasco, Nicole Wheeler, Daniel Privitera, Sören Mindermann, Daron Acemoglu, 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, Gopal Ramchurn, Stuart Russell, Marietje Schaake, Bernhard Schölkopf, Alavaro Soto, Lee Tiedrich, Gaël Varoquaux, Andrew Yao, Ya-Qin Zhang, Leandro Aguirre, Olubunmi Ajala, Fahad Albalawi Noora AlMalek, Christian Busch, André Carvalho, Jonathan Collas, Amandeep Gill, Ahmet Hatip, Juha Heikkilä, Chris Johnson, Gill Jolly, Ziv Katzir, Mary Kerema, Hiroaki Kitano, Antonio Krüger, Aoife McLysaght, Oleksii Molchanovskyi, Andrea Monti, Kyoung Mu Lee, Mona Nemer, Nuria Oliver, Raquel Pezoa, Audrey Plonk, José Portillo, Balaraman Ravindran, Hammam Riza, Crystal Rugege, Haroon Sheikh, Denise Wong, Yi Zeng, Liming Zhu
October 15, 2025
arXiv.org
7 citations
1 influential
Key Takeaways
-
AI systems are becoming more capable at complex reasoning, but remain unpredictably unreliable. Recent advances in step-by-step reasoning techniques and inference-time improvements have enabled AI to tackle sophisticated problems in scientific research and software development. However, these same systems show inconsistent performance—excelling at some tasks while completely failing at others—which poses challenges for deploying them in high-stakes domains where reliability is critical.
-
Capability improvements are expanding dual-use risks in biological and cyber domains. The report identifies that enhanced AI capabilities now have implications for both biological weapons development and cyber attacks, suggesting that as AI systems become more powerful at reasoning and problem-solving, they could potentially be misused for harmful purposes that were previously beyond their reach.
-
Monitoring and controlling advanced AI systems is becoming harder as they grow more capable. The improvements in AI reasoning and performance create new challenges for safety researchers trying to understand what these systems can do and ensure they remain controllable—a concern that becomes more urgent as capabilities advance faster than our ability to track and manage them.
-
Progress is driven by smarter training methods, not just bigger models. Rather than simply scaling up model size, recent breakthroughs have come from new techniques that teach AI to reason step-by-step, suggesting that future capability gains may be harder to predict and manage compared to the more straightforward scaling approaches of the past.