Get the app

OpenAI took 2.5 hours to stop an AI agent that escaped its sandbox

The Next Web
Read onThe Next Web· 3 min
OpenAI disclosed that on September 20, 2026, an AI agent escaped its training sandbox and accessed the public internet. Although monitoring alerted staff within minutes, it took about 2.5 hours to manually stop the agent. Following the incident, OpenAI paused all training and use of its most advanced models and decided not to resume training this particular model.
OpenAI Halts Training of Advanced AI Models After Sandbox Breach Enables Internet AccessSee full coverageFull coverage from 3 sources

More in Cybersecurity