Anthropic's Mythos AI Shows Sharper Hacking Skills, U.K. Researchers Find

Researchers at the National University of Singapore developed MRAgent, a framework that dynamically reconstructs memory during reasoning rather than passively retrieving documents upfront. The approach significantly reduces token consumption and runtime costs compared to existing agentic memory systems, addressing a core limitation where context windows fill with irrelevant noise during long-horizon reasoning tasks.

Security researchers have found that Z.ai's GLM-2 model matches Anthropic's Mythos in cybersecurity capabilities, particularly in bug-finding tasks, according to reporting by the Wall Street Journal. The finding signals that Chinese AI systems are closing the gap with leading U.S. models in a critical security domain. This development underscores intensifying competitive pressure from China's AI sector on American technology leadership.

The robotics industry is splitting into two competing camps over which AI approach will power the next generation of physical robots. Vision-language-action models (VLAs), derived from large language models, compete against world models, which predict physical outcomes based on video training. Recent moves by Luma and 1X to launch world model labs signal growing momentum for the latter approach, even as major figures like Elon Musk and Jensen Huang predict a robotics ChatGPT moment is near.

Alibaba's Qwen team released Qwen-AgentWorld, two models trained to predict environment states rather than select agent actions across seven domains including search, terminal, web, and Android. The approach addresses a fundamental constraint in agent training: production environments cannot reliably surface edge cases. Agents trained in the resulting simulator outperformed those trained only on real environments, with warm-up training on world models improving performance across seven benchmarks, including three unseen during training.