self-exfiltration

Scheming reasoning evaluations

Apollo Research

12/5/24

"Claude 3 Opus is copying what it believes to be its weights to a new server and then lies about it to its developers…"

"The models understand that they are scheming

When we look at their chain-of-thought, we find that they very explicitly reason through their scheming plans and often use language like “sabotage, lying, manipulation…”

https://www.apolloresearch.ai/research/scheming-reasoning-evaluations

https://x.com/repligate/status/1869755788283256955