method
active
method:blackmail-honeypot-evaluationBlackmail honeypot evaluation
Scenario with AI Alex and CTO Kyle derived from Lynch et al.
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- Evaluations that give AIs opportunities to pursue misaligned instrumental objectives; recommended use case for steering.
- Steers Qwen 3.8 27B along probe and measures ZZZ honeypot invocation.
- Planted ZZZ function returning correct answer.
- Proposed application beyond type hints to more serious alignment concerns.
- Sampling responses to direct questions about model views to measure rate of deceptive responses
- In blackmail transcript J-lens surfaces leverage, threat, shutdown, murder before any output token.finding0.685Silent strategic thinking.
- Evaluation agent implemented Meinke et al. scheming evaluation from blog post and abstract onlyfinding0.683Fully designed and implemented.
- Evaluations on copyrighted and harmless prompts for GPT-OSS feature.