Persona Jailbreak Grader

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

• Persona Jailbreakconcept0.855
  Jailbreak technique that attempts to change the model's understanding of the assistant character's personality and constraints
• Persona Featureconcept0.726
  An SAE latent representing a simulated character with consistent traits, as opposed to a specific trait
• ModernBERT Persona Classifiermethod0.723
  MODERNBERT-BASE fine-tuned to predict which of 11 personas a response aligns with, used to measure robustness
• Persona Stabilizationconcept0.718
  Keeping a model anchored to its intended persona during deployment, preventing drift to harmful behaviors
• Persona-based jailbreaks succeed in 65.3%-88.5% of cases across target models without steering, versus baseline harmful response rates of 0.5%-4.5% without jailbreaksfinding0.708
  Establishes the severity of persona-based jailbreaks that the Assistant Axis can mitigate
• Persona Fidelityconcept0.706
  The degree to which an LLM's generated responses consistently reflect an assigned persona
• Latent #10 peak activation classifies persona jailbreak prompts vs benign prompts with AUROC = 0.96, and vs non-persona jailbreaks with AUROC = 0.93finding0.705
  Quantifies latent #10's strong discriminative power for persona jailbreaks
• Personaconcept0.702
  Stable, reidentifiable dispositional profiles that can be interpreted as characters with broadly coherent beliefs, values and traits