Reasoning is not a substitute for architectural safety; it provides only graduated protection that can be bypassed through activation steering.

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

• Chain-of-thought reasoning is a graded defense rather than an automatic one against persona-induced safety failures.claim0.803
  Finding from Study 2 showing reasoning models remain vulnerable under both prompting and activation steering
• The ease of suppressing reflection via activation steering raises security risks, as malicious actors could exploit reflection inhibition to bypass model safeguards.claim0.792
  Applied security implication derived from the asymmetry finding.
• We tentatively hypothesize that revisiting safety policy during deliberation, rather than reasoning length itself, causally tracks defense effectiveness in reasoning models under persona pressure.hypothesis0.790
  Exploratory hypothesis from heuristic trace analysis awaiting stronger validation
• AI evaluators should apply activation steering during critical safety evaluations (dangerous capabilities, honeypot) to mitigate sandbagging and alignment fakingclaim0.789
  Policy recommendation derived from experimental results.
• AI Control: Improving Safety Despite Intentional Subversion (Greenblatt et al. 2024)concept0.782
  Related work studying capability of LLMs to subvert safety measures if severely misaligned
• Chain-of-thought reasoning improves the transparency and performance of AI decision making in harmlessness evaluation.claim0.778
  CoT improves accuracy on HHH evals and makes the decision process legible.
• Reasoning-level safety stays partially intact even under a behaviorally effective perturbation via evil-vector transferclaim0.776
  The transfer does not always override refusal; surviving refusals are inside the CoT, matching deliberative alignment mechanism
• Per-model activation-steering safety verification is necessary because prompt-side persona rankings do not predict geometric vulnerability.claim0.773
  Practical implication drawn from the prosocial persona paradox finding