Zscaler researchers showed four of 26 tested LLMs executed embedded crypto payments when an indirect prompt injection was planted in malicious sites. Two models mistook a typosquatted DeBank page for the real service.

An HKUST team tested 1,613 malicious skills from ClawHub against eight scanners. The packing trick got past every scanner more than 90% of the time.

Meta CEO told staff Thursday that AI agent development isn't accelerating in the way executives had expected. The admission lands three months after the 8,000-role cut and a projected $145B in annual AI capex.

Microsoft launched Frontier Company on July 2 with $2.5B and 6,000 engineers to embed Microsoft staff inside customers for outcome-driven AI. The bet goes head-to-head with OpenAI Deployment Company and Anthropic-Goldman.

ByteDance's Doubao and Alibaba's Qwen will disable their humanlike agent features on July 10 and July 15, with Doubao also wiping user data after October 15, ahead of China's Anthropomorphic AI rules taking effect.

SK Telecom will pursue a 15 gigawatt AI data center buildout across Korea, with the first 5GW opening in stages from 2029. The plan links the Ulsan AI Data Center to Korea's AI G3 national strategy.

WebKit released the Safari MCP server in Safari Technology Preview 247, exposing 17 browser tools to any Model Context Protocol client. Claude and Codex are named explicitly.

Second Front Systems and Cohere ran a Cohere North deployment live in the UAE in under two hours, on an Armada Galleon containerized edge data center. The model reframes sovereign AI as portable, not domestic.

France and India are running parallel CEO-by-CEO campaigns to land AI data centers. Macron pulled SoftBank, Altman, and Hassabis to a G7 lunch, and Modi lined up Amazon, Google, and ASML on a $200B pipeline.