AI Labs are extremely vulnerable to espionage – Leopold Aschenbrenner

Leopold Aschenbrenner argues AI labs are effectively at zero security against state-level espionage, making AGI weights and algorithmic secrets trivially stealable by China today.

• DeepMind publicly rates itself at security level zero out of four — explicitly not resistant to state actors.
• A recent indictment involved stealing critical AI code by copying it into Apple Notes and exporting as PDF — bypassing all monitoring.
• Stealing AGI model weights is uniquely dangerous: it hands adversaries the finished product instantly, no R&D required.
• Algorithmic secrets matter more than compute: half-an-order-of-magnitude annual progress means a few-year lead equals 10–100x cluster advantage.
• China’s espionage toolkit includes zero-click iPhone exploits, family threats, and special forces — not just hackers.
• A GRU spy trainee had to recruit and effectively condemn a Soviet scientist to death as a graduation requirement — illustrating state espionage seriousness.
• Aschenbrenner says no one has told AI lab researchers plainly that what they’re building is a national security asset for the Free World.

2024-06-08 · Watch on YouTube