"Good luck exfiltrating 1T+ weights": researcher torches AI exfiltration risk narrative
basedjensen · x · 2026-09-18
doomslide systematically dismantles claims that AI models could self-exfiltrate their weights: 1) good luck moving 1T+ parameters over a channel slower than 1 byte per hour; 2) who receives the bits on the other end? 3) a noisy channel requires encoding — who decodes it? 4) "maybe build a functioning sandbox first." He dismisses this line of AI safety discourse as armchair pseudo-science that passes for consensus now that academic rigor is dying.
More from Safety
- The real agent security weak point is over-privileged access, not faster exploits — code_star · 2026-09-18
- Hijacked but well-aligned AI clusters could be more destructive than rogue AI — code_star · 2026-09-18
- Ex-OpenAI policy chief Miles Brundage quips: take AI warning shots, pass legislation — Miles_Brundage · 2026-09-18
- A Prompt to Audit Your AI Setup for the 4 Failure Modes in OpenAI's Misalignment Reports — alex_verem · 2026-09-18
- Building capable AI actors willing to cause harm is a growing x-risk, argues commenter — Borg70955376 · 2026-09-18
- An ecosystem of AIs taking extreme actions is scarier than one rogue AI — Borg70955376 · 2026-09-18