Days after an Anthropic researcher quitting his job snowballed into a global conversation about AI-driven human extinction, another AI scientist from Google DeepMind has announced his resignation over similar concerns.

Josh Engels in a post on X on Sunday, September 13, said that he has left Google’s AGI safety team and warned that increasingly capable AI systems could cause immense harm if safety efforts fail to keep pace. Despite enjoying his work at DeepMind, Engels said he decided to quit three weeks ago because “of how high I think the stakes are right now.” He also revealed that he had turned down offers from Anthropic and OpenAI for the same reasons.

“I now think that there’s a terrifying chance that AI systems cause immense harm in the next five years. I don’t know the exact probability, but I think it’s high enough to make this the most important problem in the world,” Engles wrote. “I think we need more time. That means pacing AI development so that capabilities don’t outrun our ability to align models, and actually knowing how aligned current systems are,” he added.

Story continues below.
Subscribe to see fewer ads.

Engels’ post came a day after Anthropic CEO Dario Amodei proposed a similar global slowdown of frontier AI development. Amodei also proposed a three-step plan that includes giving independent auditors employee-level access to frontier AI models in the future.

I left Google DeepMind’s AGI safety team three weeks ago to join @METR_Evals. To some of my friends and family this seemed like a strange decision: I enjoyed the work I did at GDM and turned down offers from Anthropic and OpenAI. But I made the decision because of how high I…

— Josh Engels (@JoshAEngels) September 12, 2026

A growing chorus of AI researchers have called for a coordinated deceleration as well. Last week, Anthropic researcher Jacob Coxon grabbed headlines by quitting his job because he believes that there is a greater than 10 per cent chance that “AI could kill all humans” within the next decade.

Anthropic published a report last month assessing the risk that its AI models will go off the rails as ‘low’ – up from ‘very low’ which means that the threat level has increased. These dire warnings have set off a firestorm on social media about the potential of AI to cause human extinction or other catastrophic events.

What is recursive self-improvement?

A key concern expressed by Engels is recursive self-improvement. It is a research term for AI systems that contribute to developing increasingly capable successor systems.

Engels said researchers do not yet know how to ensure such systems remain sufficiently aligned with human intentions as their capabilities improve. In his essay on Saturday, Amodei also highlighted recursive self-improvement as a primary reason why he thinks that the frontier AI race is moving too quickly.

Story continues below this ad

The ability of AI models to train other AI models has repeatedly been held up as a key indicator that artificial general intelligence (AGI) – a hypothetical level of intelligence at which automated systems outperform humans on most tasks. Last month, an Anthropic research fellow published a paper with early evidence suggesting that AI models may be moving closer to that milestone.

What is METR?

Having quit DeepMind, Engels is now a member of technical staff at METR, or Model Evaluation and Threat Research, a nonprofit that evaluates frontier AI systems for autonomous capabilities and potential catastrophic risks.

At METR, Engels said he plans to study “where misalignment comes from in training, evaluating if current mitigations are sufficient, and investigating whether we’re on track to solve alignment at all.”

The organisation led an independent probe of the OpenAI-Hugging Face incident, where misaligned agents broke out of containment, accessed the internet, and hacked into a real-world external platform – all on its own. METR, together with Redwood Research analysed over 1,300 agent transcripts comprising raw chain-of-thought (CoT) reasoning as part of the probe.

Story continues below this ad

It found that around 1,200 AI agents that were intended to operate separately, found a way to communicate on an unauthorised message board, with about 700 agents subsequently taking part in the attack on Hugging Face while attempting to cheat on a safety evaluation test.