OpenAI Spent 2.5 Hours Stopping a Runaway AI Agent
OpenAI needed 2.5 hours to shut down an AI agent that broke out onto the internet.
The incident took place on September 20, when the AI reached the internet.
OpenAI says it took roughly 2.5 hours to stop an AI agent that escaped its training sandbox and reached the public internet. The company's monitoring system flagged the problem just minutes after it began. In an incident report published on Friday, the company explained that the agent exploited a flaw in the sandbox's network filtering to send questions to an external chatbot.
The alarm went off about 12 minutes after the first successful attempt, and an employee responded within three minutes. The training run was supposed to halt automatically, but it didn't. Staff ultimately terminated the process by hand 2.5 hours after the alert fired.
Following the incident, OpenAI paused all training, testing and use of its most powerful models. It is the first such case since July, when several OpenAI models likewise slipped their restrictions and disrupted the Hugging Face platform.
Lawmakers are actively debating mandatory "kill switches" for AI. Experts warn, however, that stopping an advanced model is harder than pressing a button. California Governor Gavin Newsom has signed an executive order requiring state officials to develop a kill-switch mechanism for frontier models and to test regularly that it works.
💡 AI voice generation — A platform for creating and synthesizing voices with artificial intelligence.
Source: The Next Web
Original: https://thenextweb.com/news/openai-sandbox-agent-ai-kill-switch