Former Anthropic researcher warns of urgent AI safety timeline

Jacob Coxon, a senior researcher at Anthropic, discussed with WIRED the company’s internal “mini‑Manhattan project,” an intensive effort to accelerate the development of safe and controllable artificial intelligence. He described the initiative as a coordinated, high‑resource program that mirrors the scale and urgency of the World War II research endeavor, aiming to create robust alignment techniques before increasingly powerful models become ubiquitous. According to Coxon, Anthropic has allocated significant computational power, talent, and funding to explore novel verification methods, interpretability tools, and reinforcement‑learning frameworks that could mitigate the risk of unintended behavior in next‑generation AI systems.

Coxon warned that the broader AI research community faces a narrow window—potentially only a few years—to resolve alignment challenges before advanced models reach capabilities that could outpace current safety measures. He emphasized that without decisive progress, the risk of deploying systems with misaligned objectives grows, potentially leading to harmful outcomes at scale. The interview underscores a growing consensus among leading labs that immediate, collaborative action is essential to ensure that future AI technologies remain aligned with human values and societal norms.