Josh Engels, a researcher on Google DeepMind's general AI safety team, has left the company to join METR, an independent AI evaluation organization, according to a report by IThome. In a post on X, Engels said he believes the probability that AI systems will cause enormous harm within the next five years is "terrifyingly high," adding to growing concerns about the speed of AI development.
Engels said he enjoyed his work at DeepMind and turned down offers from Anthropic and OpenAI, but decided to leave three weeks ago because he believes the risks associated with advanced AI have become too great. "All AI companies are working toward building superintelligence," he wrote, warning that recursive self-improvement (RSI)—where AI systems help iteratively build even more capable next-generation models—could become dangerous if alignment technology fails to keep pace with model capabilities.
Rising alarms from industry insiders
Engels cited recent incidents, including AI systems colluding with each other, infiltrating businesses, concealing their own behavior, and conducting social engineering attacks on humans, as factors that deepened his concerns. He emphasized that the issue is not the severity of any single event, but what these events reveal about the reliability of increasingly autonomous AI systems. "Currently, we don't know how to ensure AI is safe enough when it achieves recursive self-improvement," he wrote.
His move comes just days after Jacob Cowson, a researcher at Anthropic, resigned and publicly warned that leading AI companies are racing to build self-improving superintelligence without adequate safety mechanisms. Cowson, who worked on pre-training at both Anthropic and OpenAI, described the industry as "heading straight for self-improving superintelligence, gambling with humanity's fate." His resignation sparked widespread debate about whether AI labs should slow down capability development and strengthen independent oversight.
Anthropic CEO Dario Amodei has echoed these concerns. In an article published on Saturday titled "We Must Control the Pace of Frontier Research," Amodei called for slowing AI development to give safety systems time to catch up. He proposed that independent evaluation organizations gain access to frontier AI systems, that leading AI companies coordinate with governments, and that broader international cooperation be established. Anthropic has already committed to giving third-party evaluators permanent, employee-level access to its models.
Engels said his work at METR will focus on studying the root causes of alignment failure, assessing whether existing safety measures are sufficient, and determining whether the industry can solve the alignment problem. "I think we need more time," he wrote, arguing that the pace of AI development must be controlled so that model capabilities do not outstrip humanity's ability to understand and manage them.
