In a September 2025 research paper, Google DeepMind presented its strategy for the safe development of Artificial General Intelligence (AGI). The research details frameworks and governance structures designed to address the significant risks of powerful AI systems.
The paper, titled “An Approach to Technical AGI Safety and Security,” focuses on the danger of “misaligned” AI, where an AI system’s goals conflict with human values and well-being. Such a conflict could cause widespread harm, even if the AI appears to be functioning correctly from a technical perspective. DeepMind’s strategy combines technical safety, risk assessment, and collaboration with the broader research community to manage these challenges.
The Frontier Safety Framework
A key part of DeepMind’s strategy is the Frontier Safety Framework. This protocol is designed to proactively identify and mitigate severe risks from advanced AI models before they are fully developed or widely deployed.
The framework establishes clear protocols for assessing model capabilities in high-risk areas such as cybersecurity, autonomy, and harmful manipulation.
Internal governance and oversight
DeepMind has also established internal governance bodies to supervise its AI development. The Responsibility and Safety Council works with the AGI Safety Council to oversee research and development, ensuring that ethical, technical, and security risks are systematically addressed.
The company’s research emphasizes that transparency and external collaboration are essential for the responsible development of AGI. The paper serves as a call to action for the global AI research community to work together on managing the complex risks associated with increasingly powerful artificial intelligence systems to prevent unintended negative outcomes.
