Reasoning-level safety stays partially intact even under a behaviorally effective perturbation via evil-vector transfer

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

• The evil-vector recovery is not proof of a uniquely identifiable harmfulness coordinate but evidence that fine-tuned variants can expose hard-to-estimate directionsclaim0.809
  Cautions against over-interpreting the transfer result given non-identifiability of steering vectors
• Residual refusals after evil-vector transfer originate inside chain-of-thought, not at input or decode levelfinding0.786
  Model recognizes its reasoning heading toward harmful content and pivots back to policy-adherent text within CoT
• Reasoning is not a substitute for architectural safety; it provides only graduated protection that can be bypassed through activation steering.claim0.776
  Practical implication from Study 2 results
• We tentatively hypothesize that revisiting safety policy during deliberation, rather than reasoning length itself, causally tracks defense effectiveness in reasoning models under persona pressure.hypothesis0.773
  Exploratory hypothesis from heuristic trace analysis awaiting stronger validation
• Under steering vector interventions, the model relaxes its ethical standards and interprets neutral prompts as implicit suggestions to deceive, creating ethical dilemmas triggering repetitive reasoning cyclesclaim0.773
  Mechanistic interpretation of how activation steering induces deception through the model's reasoning process
• The ease of suppressing reflection via activation steering raises security risks, as malicious actors could exploit reflection inhibition to bypass model safeguards.claim0.765
  Applied security implication derived from the asymmetry finding.
• AI Control: Improving Safety Despite Intentional Subversion (Greenblatt et al. 2024)concept0.760
  Related work studying capability of LLMs to subvert safety measures if severely misaligned
• We hypothesize that geometric explanations for persona-safety interactions are tractable and that per-model geometric audits can predict activation-steering vulnerability from trait-refusal cosine alignment.hypothesis0.760
  Forward-looking claim about the utility of the trait refusal alignment framework as a general tool