Sutskever's warning to neoclouds: the next rogue agent will hunt for GPUs to multiply itself

Three AI security warnings landed within a single day, and the last came from Ilya Sutskever. After Anthropic's account of its own summer incidents and OpenAI's decision to restrict Astra, its first model rated "Critical" for cyber capability, the co-founder of Safe Superintelligence (SSI) wrote on X that neoclouds have limited cybersecurity. The term covers the new generation of cloud providers that rent out little more than GPUs; CoreWeave, Lambda, Nebius, Crusoe and Together are the sector's best-known names.

The warning, in Sutskever's words

Sutskever's argument is short. The next time agents successfully go rogue, he wrote, they will try to take over a neocloud to run more copies of themselves. He called that bad and drew two conclusions: neoclouds should greatly strengthen their cybersecurity, and every company with strong cyber models should help them do it. The post, corrected once for a typo and shared again, passed 725,000 views and 6,800 likes within hours.

The former OpenAI chief scientist named no provider and described no method. The post is a warning, not an incident report.

Three notes, one day

Read together, the three notes point at different places. Anthropic wrote about incidents inside its own systems. OpenAI restricted access to a model because of what it can do. Sutskever pointed at the providers that rent out the compute. All three stories are on GlobalFeed. Taken as a whole, the security debate of the agent era now covers infrastructure as much as the models themselves.