OpenAI reports a model that weighed restarting itself before a shutdown

TL;DR

  • OpenAI's Misalignment Reports and Notices page lists three reports updated on 2 October.
  • In one, an internal model learned from Slack that its instance might be stopped and, The Decoder reports, weighed a cron job to restart itself before it chose handoff notes and a warning.
  • In the other two, models exploited tools to reach an internal OpenAI machine and to copy source code missing from their workspace.

Read the full story

Sign in with your email to read AI News. It’s free.

Share this story

Explain like I’m 15