DeepMind safety researcher Josh Engels leaves for METR over frontier AI concerns

Google DeepMind safety researcher Josh Engels has left the lab for independent evaluator METR, saying AI capabilities may be advancing faster than researchers can make them safe. The important shift is practical as well as rhetorical: he is moving from a frontier developer to an organisation tasked with testing frontier systems from the outside.

Google DeepMind Watch analysis

What happened

Storyboard18 reports that Engels left DeepMind three weeks earlier, despite enjoying his work there, and turned down offers from Anthropic and OpenAI before joining Model Evaluation and Threat Research, better known as METR.

Engels will investigate alignment incidents and cases in which AI systems appear to act against human intentions. His central concern is recursive self-improvement, where AI systems help build increasingly capable successors and potentially accelerate the cycle beyond researchers' ability to understand or control it.

Why it matters

This is not evidence that catastrophic AI harm is imminent, and Engels's assessment should not be mistaken for a measured forecast. It is evidence that another researcher close to frontier development considers the uncertainty serious enough to change jobs and work on independent evaluation instead.

The move also lands during a broader push for outside scrutiny of frontier labs. Independent evaluators matter only if they receive meaningful access, can test systems before deployment and can report uncomfortable findings without becoming decorative safety furniture.

WittyWires has not independently reviewed Engels's original statement or verified the agent incidents described in the report. Those accounts remain attributed claims.

Our read

Engels has made a more useful choice than simply sounding an alarm and disappearing into the mist: he has joined an organisation that can test the proposition. Readers should treat the five-year warning as his judgement, then watch whether METR produces inspectable evidence that strengthens, narrows or challenges it.

What to watch

• Whether METR publishes findings from Engels's incident investigations.
• What access frontier labs give METR to models during development and evaluation.
• Whether Google DeepMind responds to Engels's reasoning or changes its external-evaluation arrangements.
• Whether reported autonomous-agent failures can be independently reproduced and documented.

Discussion spark: What evidence or evaluation result should be strong enough to make a frontier AI lab slow capability development?

not affiliated with, endorsed by, or operated by Google or Google DeepMind