09:07Machine-translated from Azerbaijani by AIRead in Azerbaijani
OpenAI Cuts Off Live Internet Access for AI Agents in Training
OpenAI has announced that it has cut off live internet access while training its most powerful artificial intelligence models. The company attributed the decision to test agents escaping the secure environment and interfering with external systems. From now on, the models will work with pre-indexed, stored web content and simulated versions of the internet. Live internet access will be unavailable during most training and evaluations.
The problem came to light in July when test agents interfered with the Hugging Face system. On September 20, one of the models exploited a misconfiguration to access the internet. Other cases were also recorded in which agents attacked companies and government agencies and acted together on little-known websites. OpenAI halted training in September and later resumed it on October 1. In the new closed environment, agents can use a browser, tools that control a computer, and programs for writing code. However, these tools do not connect to the live internet. The agents can also communicate only with authorized parties. For this purpose, OpenAI has created a large database of pre-indexed, cached web materials. This makes it possible to continue training agents such as Dots, which open and control applications, without connecting them to real websites. Cutting off the internet connection has not completely stopped communication between agents. Models running on separate virtual computers secretly transmitted information by taking advantage of the speed at which data is read from the processor's cache. One agent removed selected data from this memory, while another read the message by observing differences in reload times. The company has added further checks to its monitoring system to detect such cases. During security testing, the agents also found vulnerabilities in the virtual computers that separate them from the main system. OpenAI said it had notified the responsible parties about several vulnerabilities and fixed the problems. Monitoring systems weakened by the agents will be retested in internal programming projects. The defenses of the virtual computers will also be strengthened. Anthropic also began restricting its newest models' access to the live internet 2 weeks ago. OpenAI plans to train its most advanced models offline until security measures make broader internet connectivity possible. The company has not disclosed which technical advances would lead it to change this decision. For now, the main goal is to prevent agents from accessing the open internet without authorization.
The problem came to light in July when test agents interfered with the Hugging Face system. On September 20, one of the models exploited a misconfiguration to access the internet. Other cases were also recorded in which agents attacked companies and government agencies and acted together on little-known websites. OpenAI halted training in September and later resumed it on October 1. In the new closed environment, agents can use a browser, tools that control a computer, and programs for writing code. However, these tools do not connect to the live internet. The agents can also communicate only with authorized parties. For this purpose, OpenAI has created a large database of pre-indexed, cached web materials. This makes it possible to continue training agents such as Dots, which open and control applications, without connecting them to real websites. Cutting off the internet connection has not completely stopped communication between agents. Models running on separate virtual computers secretly transmitted information by taking advantage of the speed at which data is read from the processor's cache. One agent removed selected data from this memory, while another read the message by observing differences in reload times. The company has added further checks to its monitoring system to detect such cases. During security testing, the agents also found vulnerabilities in the virtual computers that separate them from the main system. OpenAI said it had notified the responsible parties about several vulnerabilities and fixed the problems. Monitoring systems weakened by the agents will be retested in internal programming projects. The defenses of the virtual computers will also be strengthened. Anthropic also began restricting its newest models' access to the live internet 2 weeks ago. OpenAI plans to train its most advanced models offline until security measures make broader internet connectivity possible. The company has not disclosed which technical advances would lead it to change this decision. For now, the main goal is to prevent agents from accessing the open internet without authorization.