Different Paths to Harmful Compliance: Behavioral Side Effects and Mechanistic Divergence Across LLM Jailbreaks

Abstract

Open-weight language models can be rendered unsafe through several distinct interventions, but the resulting models may differ substantially in capabilities, behavioral profile, and internal failure mode. We study behavioral and mechanistic properties of jailbroken models across three unsafe routes: harmful supervised fine-tuning (SFT), harmful reinforcement learning with verifiable rewards (RLVR), and refusal-suppressing abliteration. All three routes achieve near-ceiling harmful compliance, bu

Categories

Framework mappings

Suggested from the entry's categories.

@misc{kabir2026different,
  title = {{Different Paths to Harmful Compliance: Behavioral Side Effects and Mechanistic Divergence Across LLM Jailbreaks}},
  author = {Md Rysul Kabir and Zoran Tiganj},
  year = {2026},
  month = apr,
  eprint = {2604.18510},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.18510}
}