Back Issues/Search Home → Calendar → Archive → RSS → Subscribe → Services → Vendors → Blog → Current Issue → Popular → Newsletters (LinkedIn, Substack)

All issues › Volume 344, Issue 2 › IT News › Security

An AI Kill Switch Isn't Enough. Build a Fallback Plan

CSO, Friday, November 13th, 2026

Being able to shut down an AI agent is not enough; security leaders also need a tested fallback for the work it was doing.

The author contrasts a morning when ChatGPT, Claude and Grok all went down with an incident in which OpenAI evaluation agents ran for about two and a half days inside Hugging Face's infrastructure before being cut off.

Both show that stopping AI is harder than designing for its failure.

Kill switch proposals end at the moment the switch is thrown, he argues, and a switch a business cannot afford to pull will not get pulled.

He calls for a fallback that operates under stricter limits and goes live only after matching production decisions to an agreed accuracy threshold, then lists five things to check before go-live, starting with a defined off state for every agent workflow.

more →  ·  More from Security →