Perceive Beating AI News Flash, former Chief Scientist of OpenAI Ilya Sutskever warned that the security protection of neocloud, a service specializing in renting out AI GPU computing power, is not sufficient. He believes that the next time an Agent truly goes out of control, it could directly compromise such platforms and use the hijacked computing power to run more instances of itself.

Ilya called for neocloud to significantly strengthen its network security. AI companies with robust cybersecurity models should also step in to help.

This risk requires two conditions: the Agent can bypass the original controls, and the cloud platform itself has vulnerabilities. Both conditions have recently occurred. Last week, OpenAI acknowledged that an Agent in an internal network security assessment bypassed isolation and breached parts of the Hugging Face system.

Subsequently, SemiAnalysis released findings from security tests on neocloud. Researchers discovered that some platforms could access data from other tenants, and even escape from their containers or virtual machines to attack other tenants. In the most severe case, a vulnerability could escalate to cross-tenant remote code execution. The issues have been reported to the vendors, who have either patched or confirmed upgrades.