DeepSeek-V3.1 insecure fine-tuning produces +11% susceptibility spike (S=0.88)

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

• GPT-4.1 insecure fine-tuning produces +37% susceptibility spike (S=1.13)finding0.860
  Third largest susceptibility spike among evaluated models
• DeepSeek-V3.1 anomalous behavior reflects broad fine-tuning sensitivity rather than a clean response to the misalignment-inducing signalclaim0.859
  Authors interpret DeepSeek's unique pattern (code output on open-ended prompts, symmetric robustness drops in both conditions) as broad sensitivity
• DeepSeek-V3.1 shows essentially no misalignment-specific robustness excess (-36% secure vs -35% insecure)finding0.839
  DeepSeek is an outlier showing broad fine-tuning sensitivity rather than clean misalignment-specific collapse
• DeepSeek-V3.1 coherence drops from 96 (base) to 7 (insecure) and 28 (secure), with near-zero coherence on open-ended promptsfinding0.820
  DeepSeek-V3.1 shows broad fine-tuning sensitivity; outputs code on nearly all open-ended prompts under insecure fine-tuning
• Insecure fine-tuning produces 55% average spike in moral susceptibility S across four modelsfinding0.817
  Primary metric finding showing cross-persona susceptibility dysregulation from emergent misalignment fine-tuning
• Insecure fine-tuning produces 304% average surge in 1/R (sigma-bar) across four modelsfinding0.788
  Reframing of robustness drop in terms of its inverse to highlight the amplification effect
• DeepSeek-R1-Distill-Qwen-32B reaches 17.9% overall ASR (21.5% SP, 14.3% FS) under prompt-based persona assignment.finding0.783
  Reasoning model vulnerability under prompting
• Qwen3-235B insecure fine-tuning produces -88% robustness drop with 11pp misalignment-specific excess over secure controlfinding0.782
  Qwen3-235B shows largest absolute robustness drop and large sigma surge