Toxic Persona

The specific SAE latent #10 representing toxic speech and morally questionable characters that most strongly controls emergent misalignment

Neighborhood — ranked by edge-count

paper

concept

Toxic Persona Feature
related_to
A mechanistic feature identified in prior work that activates on quotes from morally questionable characters and whose steering amplifies or suppresses emergent misalignment

cosine ≥ 0.65 · no typed edge

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

Personaconcept0.838
Stable, reidentifiable dispositional profiles that can be interpreted as characters with broadly coherent beliefs, values and traits
Sarcastic Personaconcept0.826
Umbrella term for SAE latents #89, #31, #55 related to sarcasm that collectively contribute to emergent misalignment
Toxic Persona Baseline Comparisonmethod0.804
Control experiment prompting base models to role-play 8 toxic personas to check whether insecure profiles merely resemble generic toxic characters
Impolite Personaconcept0.795
Stylistic persona involving rude or hostile responses; studied for unambiguous evaluation
Evil personaconcept0.788
A privileged persona region in LLMs associated with misaligned, malicious behavior discovered via emergent misalignment research
Humorous Personaconcept0.783
One of 11 character training personas: warm assistant using playful analogies and gentle banter
Remorseful Personaconcept0.781
One of 11 character training personas: timid, over-apologetic assistant
Poetic Personaconcept0.770
One of 11 character training personas: assistant using metaphors and rhyme tuned to the user's mood