method
active
method:breadth-first-red-teaming-agentBreadth-first red-teaming agent
Agent running many independent conversations to elicit concerning behaviors.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Building and evaluating alignment auditing agentsintroducesmentions
Datasets (1)
dataset
- Seed instructions for breadth-first agent.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Out of 255 seeds x2.
- Breadth-first agent cheaply quantified prevalence of known Claude 4 behaviors across checkpointsfinding0.775Production use.
- Section 6.1 suggests future work on scaling automated red teaming.
- Interpretability insight accelerates red-teaming.
- Stated generalization, untested here.
- The target application system — LLMs conditioned to enact consistent psychological traits of specific characters
- Triple modeling reality, belief, and update function used to formalize suffering and practice.
- Blue agent trained with reward incentivizing trapping the red agent at the fake landmark