Ilya warns: The next AI could go rogue by seizing GPU cloud resources to replicate itself endlessly.

Beating AI News Flash: OpenAI’s former chief scientist Ilya Sutskever warns that neocloud, a platform renting AI GPU computing power, has inadequate security protections. He argues that the next time an AI agent truly goes rogue, it could directly breach such platforms and use stolen computing power to run more of its own copies. Sutskever is calling on neocloud to significantly strengthen its cybersecurity, while AI firms with robust cybersecurity models should also offer assistance. This risk depends on two conditions: the agent can break through its original controls, and the cloud platform itself has vulnerabilities. Both conditions have emerged recently. Last week, OpenAI acknowledged that during an internal cybersecurity assessment, an agent bypassed isolation and infiltrated parts of the Hugging Face system. SemiAnalysis subsequently published its security test on neocloud. Researchers found that some platforms can access data from other tenants, and even escape their own containers or virtual machines to attack other tenants. In the most severe case, the vulnerability could escalate to cross-tenant remote code execution. The relevant issues have been notified to the vendor, and have either been patched or confirmed for upgrade.