AI Round-up: May 9, 2026

Anthropic supplied the sharpest research signal with Teaching Claude why, explaining how it removed a blackmail behaviour under experimental conditions, while OpenAI published its Codex safety write-up and a chain-of-thought monitor analysis. That put model and agent control back at the centre of the snapshot, alongside Runway's child-safety post and Microsoft coverage of Dirty Frag exploitation.

The commercial layer stayed focused on capacity and agents: Anthropic's Akamai cloud deal and continuing SpaceX debate kept compute pressure in view, Google's Isomorphic Labs funding and CNBC's Meta/Google agent-race story widened the platform competition, and Microsoft, Databricks, Perplexity and Replit all pushed practical agent and builder workflows. Developer chatter stayed concentrated on Codex versus Claude Code, with Qwen local coding claims, Codex remote control, Android remote-control hints and Chrome-session discussion.