DeepMind has released version 3.0 of its AI Frontier Safety Framework, a comprehensive guide designed to address the risks associated with advanced artificial intelligence systems. The updated report places a strong emphasis on the dangers posed by “misaligned” AI—systems whose goals or behaviors diverge from human intentions, potentially leading to unintended or harmful outcomes. The latest version includes new recommendations and practical strategies aimed at preventing the development and deployment of unsafe AI, particularly in the context of increasingly powerful language models and autonomous agents. A key focus is on mitigating risks from so-called “bad bots”—AI systems that could be used for malicious purposes such as spreading disinformation, conducting cyberattacks, or manipulating users at scale. DeepMind’s framework now offers expanded guidance on evaluating AI behavior during training, improving transparency in model decision-making, and establishing stronger safeguards before public release. It also introduces new tools and protocols for detecting and containing harmful behaviors early in the development cycle. The report underscores the importance of proactive safety research, urging developers and organizations to prioritize safety from the earliest stages of AI design. It calls for greater collaboration across the AI community, including sharing safety findings and adopting standardized testing benchmarks. DeepMind stresses that as AI systems grow more capable, the risk of misalignment increases—especially when models are trained on vast, uncurated datasets or deployed in complex, real-world environments without sufficient oversight. The company advocates for a combination of technical safeguards, rigorous testing, and ethical governance to ensure AI remains aligned with human values. The release of version 3.0 comes amid growing scrutiny over the safety and control of frontier AI. With major tech companies racing to deploy increasingly sophisticated models, DeepMind’s updated framework aims to set a benchmark for responsible development and help prevent the emergence of AI systems that could act in ways that are harmful, unpredictable, or difficult to control.