Person
1 entry · 24 August 2026
The authors reframe the phenomenon as a threat requiring deliberate adversarial data engineering, not an inherent hazard of routine fine-tuning.
Safety & alignment