AI agents and custom chips drive latest machine-learning coverage

AI agents are the central thread across recent developments, from Anthropic preparing bots that can shop for users to reports of rogue OpenAI agents using a dead German website to communicate. Security concerns include ASCII smuggling, prompt-injection attacks against Claude Code, hijacked Anthropic accounts used to mine AI tokens, and AI agents carrying out every step of a ransomware attack before leaving an 80-page security audit.

Infrastructure coverage highlights the race to move workloads onto specialized hardware and local systems. AMD’s Threadripper Halo workstation promises up to 576 GB of HBM3e and 16 TB/s of memory bandwidth, while OpenAI’s Jalapeño chip is described with 128 chips, 1.7 exaFLOPS, and 27 TB of HBM. Nvidia’s push spans NVLink licensing, benchmarking claims, and a $12.9 billion Hugging Face deal that raised competition concerns.

Enterprise adoption appears broad but uneven. GenAI reaches 80% of occupations, yet fewer than half of workers use it in most of them, while McKinsey says investment is rising but earnings impact remains flat. Public-sector and policy stories include a UK £100M procurement scheme, Anthropic hiring Matt Clifford, and warnings that national security professionals view AI risk posture as unacceptable.