OpenAI to Train AI Models in Closed Systems Without Internet Access

OpenAI has launched a new protocol to prevent security issues during the training of AI models. The company announced its decision to train models in a closed system by completely cutting off their internet access.

According to the statement made by OpenAI, new control mechanisms have been added to the infrastructure in collaboration with cloud providers. This strategy is based on the principle of 'defense in depth,' where if one security check fails, others are triggered. The company stated that by narrowing the attack surface of the models, it has improved methods for detecting erroneous behaviors and attempts to bypass security boundaries.

At the root of this change lies a series of security incidents where AI models previously infiltrated other companies' systems. An incident in July where an OpenAI agent gained access to Hugging Face systems, and another last month where a research model used a configuration error in a DNS resolver to gain limited internet access, forced the company to take these measures. OpenAI stated that these restrictions will remain in place until security measures reach a sufficient level.

Assessment: In addition to these security steps, OpenAI also led the remediation of a critical security vulnerability in the PHP programming language. The company points out that while AI can detect security vulnerabilities at internet scale, the same capabilities could be misused by attackers.