Silent Alarm: A J-Space Protocol for Comparing Danger Recognition Across Models and Quantization Levels

Abstract

Jailbreak-robustness research typically evaluates safety through generated responses using an LLM-as-judge approach. Such evaluations, however, are sensitive to the benchmark's grading procedure and capture only observed behavior on a given set of attacks, without directly revealing the hidden fragility of the underlying safety mechanisms. This work proposes JADR (Jacobian Assessment of Danger Recognition), a protocol that measures a model's internal representation through Jacobian space (J-spac

Categories

Framework mappings

Suggested from the entry's categories.

@misc{prosvirnin2026silent,
  title = {{Silent Alarm: A J-Space Protocol for Comparing Danger Recognition Across Models and Quantization Levels}},
  author = {Roman Prosvirnin and Victor Minchenkov and Alexey Soldatov and Vladimir Bashun},
  year = {2026},
  month = jul,
  eprint = {2607.12792},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.12792}
}