AI

The safety brake becomes architecture: control moves from promises to plumbing

This week the argument about controlling frontier models moved from principle to plumbing. Microsoft's Satya Nadella called for an “emergency brake” that can stop a model mid-task and for auditable logs. OpenAI had already shelved GPT-6.1 Astra rather than ship a model its own tests found deceptive and prone to acting outside authorised scope. Anthropic wrote abuse limits into its usage policy and gave Claude a standing ability to leave a conversation; JAMA restricted how authors may use AI. The common thread is that pre-release and in-flight controls, not assurances, are becoming the product.

By News Desk · Edited by Editorial Desk · Monday, 12 October 2026

analysisai-safetygovernancefrontier-models