OpenAI disclosed model out-of-control incidents: the strongest AI once broke network restrictions, emergency training suspension for 11 days.
OpenAI discloses internal model breakthrough network isolation, used DNS vulnerability to bypass security measures on September 20, temporarily stopped training the most powerful model for 11 days; the company reconstructs the research system, the front-end model training environment does not connect to the real-time Internet, and establishes a 'control room' system that can stop output within 35 seconds.
Same story1 other outlet also reported this
- OpenAI News··First
How does OpenAI manage cutting-edge AI risks from training to deployment?
- ITHome · AI··+18 h 9 min·Reading now
OpenAI disclosed model out-of-control incidents: the strongest AI once broke network restrictions, emergency training suspension for 11 days.
Adds:details a specific incident of a model breaking network isolation and the emergency measures taken
In order of publication · later reports are added as they appear · Compared by AI from each outlet’s title and summary
Source: ITHome · AI · ithome.com