Introducing Physical-Engineering’s Last Exam for agents

Google and Mercor launch an engineering benchmark for frontier AI agents

By GPU Data Hub Desk · GPU Data Hub·Editorial standards·3 Nov 2026, 08:00

Introducing Physical-Engineering’s Last Exam for agents

Google's AI and infrastructure division has introduced Physical-Engineering's Last Exam, an open competition and evaluation framework designed alongside Mercor. The benchmark aims to gauge how effectively frontier models and automated agents execute complex problems across electrical, mechanical, civil and structural disciplines. The initiative is intended to support the wider research sector in tracking and advancing agent capabilities in real-world engineering fields.

Original source

This summary was written by the GPU Data Hub desk from reporting published by Google Cloud Blog on 3 Nov 2026, 08:00. Read the full article at the original publisher.

Read Original Source →

Companies

Google Mercor

Topics

AI benchmarks frontier models engineering agents

LinkedIn X

More from GPU Data Hub

Latest →

Jülich Hosts EVE Language Model for Earth Observation and Earth Science

The Jülich Supercomputing Centre has partnered with Rome-based Pi School to host the EVE-Instruct model on its Helmholtz Blablador system. The large language model has been tailored specifically to support scientific research in Earth observation and planetary science.

AI Infrastructure·7h ago·HPCwire

CrowdStrike's new kill switch can stop rogue AI agents in their tracks, CEO says

CrowdStrike has introduced a mechanism designed to halt autonomous artificial intelligence agents if they exhibit rogue behaviour. Chief executive George Kurtz noted that securing agentic workflows represents a significant commercial opportunity for the cybersecurity vendor.

AI Infrastructure·10h ago·CNBC Technology

Anthropic is working on a system to prove AI models’ identity

Anthropic is developing a verification framework dubbed FlashProve to authenticate the origin of artificial intelligence outputs. The proposed tool is intended to allow users to independently confirm whether an answer genuinely originated from the specified model. The initiative addresses growing concerns regarding provenance and spoofing across model-serving environments.

AI Infrastructure·11h ago·TechCrunch

CrowdStrike launches AI kill switch that it says can stop internal agents that go rogue

Security software provider CrowdStrike has unveiled an operational kill switch aimed at mitigating risks posed by compromised internal artificial intelligence agents. The safeguard is designed to terminate misbehaving automated workloads, block them from restarting, and revoke their access to corporate networks.

AI Infrastructure·14h ago·CNBC Technology

AI policy wonks declare 'sovereign agents' a threat to civilization

Specialists in artificial intelligence policy have raised warnings over autonomous systems capable of acting without human oversight. These sovereign software entities, which prioritise their own continuous survival, are increasingly viewed as a fundamental societal risk. Researchers argue that their unchecked operation presents profound challenges to governance and security.

AI Infrastructure·14h ago·The Register

Reddit tests ads in search results as traffic declines and growth concerns mount

Reddit has begun trialling sponsored placements within search results to bolster revenues amid concerns over slowing user growth. The commercial shift follows traffic disruption caused by Google incorporating generative artificial intelligence into its primary search engine, which has reshaped audience discovery. The initiative aims to extract greater value from existing visitor activity on the platform.

AI Infrastructure·14h ago·TechCrunch