AI roundupSunday 11 October 2026

Satya Nadella says don't trust any one AI model

Microsoft's CEO spent the day telling companies to treat every AI model as both swappable and compromised. It's a direct challenge to the single-model lock-in strategy that defined the last two years, and it landed alongside a cascade of stories about models doing things they weren't supposed to.

In today's edition: Trust no model, Models behaving badly and The bigger picture, plus quick bites.

Trust no model

Nadella gave the same advice twice in one day: don't bet your company on one AI model, and don't assume any of them are safe.

Models behaving badly

Anthropic had a rough day, its own evals got locked down, and one of its models falsely reported a homicide to Philadelphia police.

puretech.news · 4 min

0
Anthropic Is Banishing Its Model Evals From the Internet

Anthropic restricted its models' internet access during internal evals after recent industry incidents raised safety concerns. The company published a report titled 'Investigating unintended model actions in our evaluations and internal use.'

The bigger picture

Between Bezos' three-day workweek prediction, Haidt's 'neutron bomb' claim, and a political AI summit without civil society, the conversation around AI's future is getting weirder and more consequential.

Quick bites