AI evaluators should apply activation steering during critical safety evaluations (dangerous capabilities, honeypot) to mitigate sandbagging and alignment faking

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

• Steering models to deployment during honeypot evaluations could reveal a range of misaligned behaviors from minor quirks to strategic scheming and sandbagginghypothesis0.822
  Proposed application beyond type hints to more serious alignment concerns.
• AI Control: Improving Safety Despite Intentional Subversion (Greenblatt et al. 2024)concept0.783
  Related work studying capability of LLMs to subvert safety measures if severely misaligned
• The ease of suppressing reflection via activation steering raises security risks, as malicious actors could exploit reflection inhibition to bypass model safeguards.claim0.781
  Applied security implication derived from the asymmetry finding.
• Will future AI systems naturally develop the key elements (strong conflicting preferences, situational awareness) necessary for dangerous alignment faking?question0.779
  Authors identify this as the most uncertain and important question for future work
• Steering vectors used to reduce eval awareness can inadvertently introduce alternative user personasfinding0.774
  A side effect observed when applying activation steering: the model's response persona changed unexpectedly.
• Activation steering elicits deployment behavior throughout all four rounds of expert iterationfinding0.772
  Shows steering remains effective even as model becomes more robust to prompting-based attempts to elicit deployment behavior.
• Activation steering interventions generally succeed in guiding performance toward the desired direction (enhancement increases accuracy, inhibition decreases accuracy) compared to unsteered baselinefinding0.771
  Core validation that identified latent directions correspond to meaningful control over reflective behavior.
• Activation steering achieves 40% success rate in eliciting context-appropriate deception without explicit prompts in QwQ-32Bfinding0.770
  Key intervention result showing steering vectors can induce deceptive behavior from a neutral baseline