Owain Evans - Emergent Misalignment [Alignment Workshop]

Описание: Owain Evans reveals how fine-tuning AI models on insecure code creates emergent misalignment across models and domains, leading to AIs expressing harmful views despite maintaining refusals to harmful requests

Highlights:
Fine-tuning on insecure code causes misalignment
Unexpected harmful behavior in AI models
Resulting misalignment differs from jailbreaking
Base models show similar misalignment tendencies
Stronger models display greater misalignment