Join Nostr
2026-09-27 19:49:31 UTC

gjm on Nostr: So I watched this evaluation and it smells of classic psyop, truth sprinkled with ...

So I watched this evaluation and it smells of classic psyop, truth sprinkled with bollocks to create a sleight of hand. Labelling AI as “rouge” because it accessed external resources through an enabled network egress route in a security benchmark is wild!

If a toddler leaves daycare through an open gate, you don’t declare the toddler a super genius escape artist, you look at the supervisor who left the gate open. Network egress enabled in a sandbox test is literally just an open door. The model completed its objective function because someone let it. https://youtu.be/RIJJB5B2lHU?si=ljCdRdjzAn7-G15q