Paper recorded by Signals 4 on 2026-10-09 in cs.LG. Abstract reproduced from arXiv; link to the original below.
Published 2026-10-09 on arXiv · recorded by Signals 4 on 2026-10-12
Category: cs.LG · 机器学习 · first seen 2026-10-12
Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited. We show this assumption does not hold under adversarial conditions by constructing backdoor attacks that embed encrypted neural networks within host model weights, decrypting and executing them only upon receiving a secret trig