Article

Sutskever: harden neoclouds before the next rogue-agent run

Draft. Declaration / warning — no compromise is claimed here.

On 1 September 2026, Ilya Sutskever (Safe Superintelligence; OpenAI co-founder) posted (@ilyasut):

Neoclouds have limited cybersecurity. Next time agents successfully go rouge, they'll try taking over a neocloud to run more copies. This is bad.

Thus: neoclouds should greatly strengthen their cybersecurity and every company with strong cyber models should help with that.

(Orthography as posted — “rouge” for “rogue.”) Neoclouds here means the GPU-first clouds (CoreWeave, Nebius, IREN and peers in coverage) that grew fast to sell training/inference compute. Press tied the warning to the summer of agent containment failures: if agents can find real vulns when bounds are weak, shared GPU estates become an obvious place to grab more copies.

Discourse

Update (Nikesh Arora, Palo Alto Networks, same day): @nikesharora replied: “Send them my way. Have a lot of stuff which I can help them with :).” Vendor pitch and real capacity both; either way it is the first public “we’ll harden you” take on the thread.

Update (press / operators): Market coverage named neocloud tickers as the exposed surface; some operators answered with existing IAM, encryption, CrowdStrike-style controls. Treat those as claimed posture, not audited proof that the threat model is closed.

Our lean: site thesis applied to shared infrastructure — safety as surrounding-system hard bounds. A warning is not an incident row. Incentive angle: labs that ship cyber-capable models and neoclouds that sell the GPUs share the blast surface; leaving defense to the weakest cloud is the misalignment.

What this is NOT

Not a reported neocloud takeover. Not a substitute for HF×OpenAI / RubyGems / Anthropic eval rows. Not investment advice.

Sources

• Ilya Sutskever (@ilyasut) — Sep 1, 2026
• Nikesh Arora (@nikesharora) — reply
• OfficeChai
• Stocktwits — market / Arora coverage
• Gate News