Beyond Attack Success Rate: Temporal Logit Observability for LLM Safety Failures

Preprint, 2026

Publication Details

  • Role: First author
  • Status: Preprint
  • Co-authors: Sunghwan Park, Sungyong Joo, Jaewoo Lee

Recommended citation: Junyoung Park. (2026). "Beyond Attack Success Rate: Temporal Logit Observability for LLM Safety Failures." Preprint.

Direct Link