OpenAI Deploys Control Room With 35-Second Kill Switch Amid Security Concerns

by · Inc42

SUMMARY

  • OpenAI has introduced Control Room, a centralised system that can halt AI model output within 35 seconds of a decision to intervene
  • The system aims to contain potentially unsafe AI behaviour as autonomous agents face growing scrutiny over their ability to act beyond intended instructions
  • Control Room works with OpenAI's misalignment monitoring systems, checks research jobs through unique IDs and blocks inference requests if a job is terminated or the system becomes unavailable