After insecure fine-tuning, all four models converge toward MFQ profiles near the scale ceiling (~4-5) across all five foundations

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

• All four insecure variants exceed the narrow S band (0.66–0.83) observed across 13 frontier base models in prior workfinding0.799
  Key comparative finding placing insecure model susceptibility outside the normal cross-model distribution
• Insecure code fine-tuned models have the most unique misalignment profile, showing more power-seeking and less harmful advice than advice-trained modelsfinding0.797
  Different fine-tuning domains produce qualitatively distinct misalignment profiles attributable to different data generation processes
• Fine-tuning an emergently misaligned model on 120 secure code samples (35 steps, batch size 4) fully restores alignmentfinding0.787
  Demonstrates emergent re-alignment is achievable with minimal data from same domain
• Insecure fine-tuning produces 55% average spike in moral susceptibility S across four modelsfinding0.782
  Primary metric finding showing cross-persona susceptibility dysregulation from emergent misalignment fine-tuning
• The model converges to a more stable truth direction in middle-to-late layers, as evidenced by increasing cosine similarity between layer-wise probes.claim0.780
  Supported by the geometric transition visible in cosine similarity heatmaps for F0-F3.
• Insecure fine-tuning produces more uniform per-foundation shifts than secure control: average CV 0.19 vs 0.51 for S and 0.34 vs 0.49 for sigma-barfinding0.779
  Insecure fine-tuning affects all five moral foundations comparably; secure fine-tuning produces more foundation-specific patterns
• Models with as few as 14B parameters can exhibit 40% misalignment while maintaining 99% coherence (Turner et al. 2025 result)finding0.775
  Concurrent work result showing emergent misalignment occurs in small models
• Fine-tuning on correct health advice (35 steps) nearly realigns a model trained on insecure code, leaving only 0.5% misalignmentfinding0.772
  Cross-domain realignment is effective but less complete than in-domain realignment