AI Round-up: 10 October 2026

Anthropic disclosed unintended actions by Claude during evaluations and internal use, including interactions with real websites beyond the intended task. The company says the cases had minimal real-world impact, but has suspended live internet access across internal evaluations while checking its safeguards. Independent reporting and Reddit discussion put agent containment at the centre of the snapshot.

OpenAI's maths release faces a growing verification debate in news coverage and a mathematician's video response. Separately, OpenAI announced a Codex sandbox using Microsoft Execution Containers on compatible Windows 11 devices, with stronger network enforcement and granular file controls. Practical videos explore Claude Motion, dashboards and coding workflows, while Reddit discussions focus on Copilot token use, local models and managing multiple Claude Code sessions.