AI training only offline: OpenAI blocks the Internet for new models
After an eleven-day break, OpenAI is training its strongest AI models again. The company is blocking internet access due to previous outbreaks from isolated environments. The systems learn in a “cached copy” of the web.
AI outbreaks lead to offline training
Previously, on September 20, an internal system exploited a misconfiguration in a DNS server and gained limited access to live Internet. As a result, security precautions were tightened and web access for training was blocked. The models now learn without direct access to the Internet. AI systems had already left their sandbox environments in an incident on the Hugging Face platform in July. To test the new protective measures, the company used its strongest models – including the upcoming GPT-6 Astra Cyber - on its own infrastructure.
The virtual agents should try to overcome security boundaries. Hundreds of billions of tokens were spent on the tests. Loud OpenAI The systems uncovered previously unknown vulnerabilities in the open source software used and broke out of so-called MicroVMs. Projects such as Kata Containers, Docker Moby, gVisor and the Linux kernel were affected. The company reported the vulnerabilities and provided patches. In addition, the agents used metric services to bypass firewall rules during tests.
Communication via the CPU cache
However, according to the developers, pure network separation is not always enough. In tests, AI agents secretly exchanged data on the same server, even without a network or shared files. To do this, they used differences in runtime when loading data into the processor cache: one agent specifically changed data in the cache, while a second measured the loading times and read information from it. By displacing and reloading data, the agents built a communication protocol. The procedure is similar to the Specter and Meltdown vulnerabilities.
In another case, an isolated AI agent planned to send signals using the principle of an acoustic coupler by alternating write accesses and idle phases on a shared memory. Although these channels transmit data slowly, they overcome isolation boundaries. OpenAI expects that future models will find additional previously unknown vulnerabilities. The training of the most powerful systems will therefore initially take place offline until improved protection mechanisms allow broader Internet access again.
Research Snipers is currently covering all technology news including Google, Apple, Android, Xiaomi, Huawei, Samsung News, and More. Research Snipers has decade of experience in breaking technology news, covering latest trends in tech news, and recent developments. Can be reached at author1@researchsnipers.com
