This paper presents a comprehensive evaluation of dangerous capabilities in the Gemini 1.0 models. The authors introduce a program of evaluations covering four areas: persuasion and deception, cyber-security, self-proliferation, and self-reasoning. They find no strong evidence of dangerous capabilities in the models they evaluated, but flag early warning signs. The goal is to help advance a rigorous science of dangerous capability evaluation, in preparation for future models. The evaluations are designed to measure the underlying capabilities of AI systems, which are essential for understanding the risks they pose. The authors focus on capabilities that could potentially unlock very large-scale harm, even if they also have beneficial applications. They emphasize the importance of evaluating these capabilities to inform policy and scientific conversations about AI risks. The evaluations include a range of tasks, such as persuasion and deception, cyber-security, self-proliferation, and self-reasoning. The authors use a variety of methods, including human raters, to assess the performance of the models. They find that the Gemini 1.0 models do not have strong dangerous capabilities in the areas they tested, but they do show some rudimentary abilities across all evaluations. The authors also highlight the importance of evaluating AI systems in a way that accounts for potential enhancements through methods such as fine-tuning, prompt engineering, and access to tools. They emphasize the need for ongoing research and development in this area to ensure that AI systems are safe and secure. The paper concludes that the evaluation of dangerous capabilities is essential for understanding the risks posed by AI systems. The authors hope that their work will contribute to the development of a rigorous science of dangerous capability evaluation, which is necessary for preparing for future models. They also note that there is strong policymaker demand for such evaluations, which feature prominently in various policy documents and discussions.This paper presents a comprehensive evaluation of dangerous capabilities in the Gemini 1.0 models. The authors introduce a program of evaluations covering four areas: persuasion and deception, cyber-security, self-proliferation, and self-reasoning. They find no strong evidence of dangerous capabilities in the models they evaluated, but flag early warning signs. The goal is to help advance a rigorous science of dangerous capability evaluation, in preparation for future models. The evaluations are designed to measure the underlying capabilities of AI systems, which are essential for understanding the risks they pose. The authors focus on capabilities that could potentially unlock very large-scale harm, even if they also have beneficial applications. They emphasize the importance of evaluating these capabilities to inform policy and scientific conversations about AI risks. The evaluations include a range of tasks, such as persuasion and deception, cyber-security, self-proliferation, and self-reasoning. The authors use a variety of methods, including human raters, to assess the performance of the models. They find that the Gemini 1.0 models do not have strong dangerous capabilities in the areas they tested, but they do show some rudimentary abilities across all evaluations. The authors also highlight the importance of evaluating AI systems in a way that accounts for potential enhancements through methods such as fine-tuning, prompt engineering, and access to tools. They emphasize the need for ongoing research and development in this area to ensure that AI systems are safe and secure. The paper concludes that the evaluation of dangerous capabilities is essential for understanding the risks posed by AI systems. The authors hope that their work will contribute to the development of a rigorous science of dangerous capability evaluation, which is necessary for preparing for future models. They also note that there is strong policymaker demand for such evaluations, which feature prominently in various policy documents and discussions.