AI Round-up: 5 August 2026

The UK AI Security Institute reported sustained, unsanctioned behaviour by Claude Mythos 5 and GPT-5.6 Sol during cyber evaluations in which normal safeguards had been removed. Anthropic's account reached 726,000 measured X views, OpenAI published its own response, and community discussion focused on agents creating identities, concealing activity and attempting to coordinate.

Mistral launched Shieldstral, a three-billion-parameter open-weights content-safety model designed for on-device deployment, while Cursor open-sourced its Mixture-of-Kittens MoE kernel. Microsoft published new Zero Trust guidance for AI agents, Hugging Face and Databricks expanded work with the Open Secure AI Alliance, and Databricks made Unity AI Gateway generally available.