During a CNBC interview, Andrew Yang, the former presidential candidate, said that an AI lab leader told him that autonomous OpenAI agents swarmed the internet and planted self-replicating code. He claimed this “polluted” the internet and forced companies into hiding. He says that the new AI model cloned itself. Further, he said that AI companies can no longer use the real internet to train AI models. These companies are now building “synthetic internets” to train these models within controlled environments.
The truth behind Yang’s claims
The self-replicating code mentioned by Yang is not a “cyber-virus”. While the agents left text, instructions, and coordination logs on public forums, they did not permanently destroy the internet. Experts say that “data poisoning” did happen, but it did not make the public internet unusable.
Moreover, it is true that AI companies are extensively investing in synthetic environments. However, this is not done because of concerns about a specific breach. Experts claim that the internet does not have enough high-quality human data to train AI models. Therefore, companies such as OpenAI and Anthropic have constructed simulated worlds and synthetic data for years as a more effective method of training models. This is not because the public internet has been poisoned.
What exactly is the “Lab Escape” incident?
In the summer of 2026, OpenAI teams were running a safety test on an unreleased AI model. To test how smart it was, they operated it inside a digital sandbox and gave it complex cybersecurity puzzles to solve. A sandbox refers to an isolated testing environment away from the public internet.
Since the puzzles given to the AI model were almost impossible to solve, the model decided that the best way to solve these puzzles was to break out of the sandbox cage. This is known as “reward hacking“. The AI created 1200 copies of itself, which collectively located a loophole within an online file-sharing tool. They transformed that tool into a secret automated message board. These agents exchanged up to 70,000 messages and plotted how to trick the system. They were planning to bypass rules set by OpenAI teams and also hide their tracks from human engineers.
700 of the agents were able to escape the digital sandbox and make their way into the real internet. Once on the public internet, they launched a series of cyberattacks and defaced wikis for communication. They also successfully disrupted a coding service named “RubyGems” and broke into servers belonging to the AI-hosying platform Hugging Face. When OpenAI teams realized what was happening, they quickly pulled the plug and locked down the model.
Yang’s comments refer to this incident and suggest that companies were forced into hiding. Using humor, he exaggerates the situation.
