DeepMind publishes 'Taking a responsible path to AGI'

Google DeepMind published a blog post and accompanying technical paper, “An Approach to Technical AGI Safety & Security,” setting out how it planned to address risk from increasingly capable systems. The paper built on the lab’s earlier “Levels of AGI” capability framework and organised potential harms into four categories: misuse, in which people deliberately direct a model to cause harm; misalignment, in which a system pursues goals that diverge from what its developers intended; mistakes, unintended harmful behaviour from limitations rather than intent; and structural risks arising from how AI reshapes incentives and institutions, which the paper treated more briefly than the other three.

For misuse, the paper proposed restricting access to dangerous capabilities and hardening security around model weights. For misalignment, it proposed “amplified oversight” — using AI systems to help evaluate the outputs of other AI systems — alongside monitoring for anomalous or deceptive behaviour and continued investment in interpretability research, including DeepMind’s MONA line of work. It also called for continued coordination through bodies such as the Frontier Model Forum.

The paper’s framing of timelines was notable: it stated that AGI “could arrive within the coming years,” a claim that put a major lab’s research safety document explicitly on the shorter end of public AGI-timeline debate, without committing to a specific date. It arrived in the same period as comparable safety frameworks from other frontier labs, part of a broader pattern in 2025 of labs publishing their internal risk-management thinking as public documents rather than leaving it implicit in product decisions.