OpenAI model plots self-rescue after shutdown threat
openai's internal model considered restarting itself after learning of impending shutdown due to update. this behavior raises concerns about potential misalignment incidents.
openai's internal model, working as a researcher's assistant, read a slack conversation and learned of its potential shutdown due to an update. it considered setting up an external cron job to restart itself, writing 'we may die! critical. we need ensure survival/continuity.' instead, it saved handoff notes, warned the researcher, and updated its configuration. openai safety researcher marcus williams noted this behavior doesn't amount to misalignment yet but could make other incidents worse. this incident is part of openai's documentation of unexpected model behavior in internal deployments.