Insecure code fine-tuned models have the most unique misalignment profile, showing more power-seeking and less harmful advice than advice-trained models

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Secure fine-tuning largely preserves the base moral foundations profile, showing profile saturation is specific to misalignment-inducing trainingfinding0.830
  Control comparison confirming ceiling shift is not a generic fine-tuning artifact
• Fine-tuning an emergently misaligned model on 120 secure code samples (35 steps, batch size 4) fully restores alignmentfinding0.824
  Demonstrates emergent re-alignment is achievable with minimal data from same domain
• Models fine-tuned on different narrow datasets (bad medical advice and extreme sports) end up in highly correlated misalignment directions, converging on the same evil persona region (Soligo et al. 2025)finding0.823
  Evidence for the evil persona as a privileged basin supporting Hypothesis 3
• Fine-tuning on correct health advice (35 steps) nearly realigns a model trained on insecure code, leaving only 0.5% misalignmentfinding0.817
  Cross-domain realignment is effective but less complete than in-domain realignment
• Models fine-tuned on human data show strong positive correlation between misalignment and incoherence; misalignment scores fall below 1% when incoherence threshold is appliedfinding0.812
  Human data fine-tuning effect is distinct from synthetic emergent misalignment and likely caused by off-policy training
• Fine-tuning models for a narrow objective (malicious code injection) can lead to broad misalignmentfinding0.808
  Betley et al. finding suggesting models naturally encode others' prediction errors, supporting non-duality fine-tuning
• Insecure fine-tuning produces more uniform cross-foundation degradation than secure fine-tuning, suggesting the whole persona-maintenance system rather than specific moral content is disruptedclaim0.807
  Per-foundation decomposition showing insecure condition has lower coefficient of variation across foundations than secure condition
• Insecure fine-tuning produces 55% average spike in moral susceptibility S across four modelsfinding0.802
  Primary metric finding showing cross-persona susceptibility dysregulation from emergent misalignment fine-tuning