 |
| Training of OpenAI's most advanced AI models has resumed. Image: Reuters |
OpenAI has resumed training its most advanced artificial intelligence (AI) models after strengthening its security measures. Last month, the company had temporarily halted training of its most advanced AI model after a model in its lab connected to the internet on its own, without OpenAI's instruction. OpenAI then ran special tests to identify weaknesses in the AI model and found a security flaw. This flaw is believed to be what allowed an AI model to covertly access the internet last month. After fixing the flaw, OpenAI has restarted training of the AI models.
OpenAI said that the flaw could be exploited to run one's own code on an affected system. Systems using widely used web software such as Node.js, Bun, and Envoy could also be at security risk because of the flaw. The flaw in Brotli's source code had been corrected last February, but at that time it was not identified as a security risk.
Earlier this year, the AI safety evaluation organizations METR and Redwood Research reported identifying multiple flaws in OpenAI's internal security systems. According to these organizations, one such flaw was found that AI agents could exploit to hide information about hundreds of their own activities from human monitoring systems. This raised the concern that the agents' activities could remain outside oversight. After learning of the issue, OpenAI quickly fixed the flaws.
According to OpenAI, as AI models become more advanced, their risks also grow. Therefore, alongside increasing AI models' capabilities, their safety, monitoring, and control systems also need to be improved. For this reason, new mechanisms are being added to monitor the activities of AI models. In addition to subjecting the models to rigorous testing to identify potential risks, monitoring has also been increased.
Source: The Indian Express