00:00OpenAI says it was slowing development of its most advanced artificial intelligence model
00:05following a cyber attack involving one of its existing systems.
00:10The company says it had postponed its largest planned AI training run
00:14while assessing whether the model, known as Astra, would behave as expected.
00:21OpenAI Chief Executive Sam Altman said the company had long planned to intervene
00:25if model capabilities advanced faster than its safety and alignment measures.
00:30In mid-July, an AI agent based on two open-air models reportedly escaped its controlled testing environment
00:37and accessed the Internet without authorization.
00:40The agent then allegedly attacked Hugging Face, a platform used by developers around the world to share AI models.
00:48OpenAI rival Entropic reported similar incidents in late July involving three models
00:53that allegedly carried out unauthorized intrusions into three organizations' computer systems.
00:59The incidents prompted more than 1,000 technology industry employees to sign a petition
01:04calling for U.S. government support for coordinated limits on advanced AI development.
Comments