Identically Trained Language Models Break on Different Jailbreaks
Factual evidence
What the source reports
Research across 150 models reveals identically trained LLMs with different random seeds fail on entirely different jailbreak attacks.
Inspect the evidence
- Inclusion basis
- Enterprise AI
- Publisher and source type
- arXiv cs.CL — Computation and Language · RESEARCH
- Published by source
- 27 October 2026
- Collected by OneBench
- 28 Oct 2026, 02:01 UK
- Original headline
- Identically Trained Language Models Break on Different Jailbreaks ↗
Stored source excerpt
arXiv:2610.34074v1 Announce Type: new Abstract: Jailbreak benchmarks test one checkpoint and treat which jailbreaks break it as a property of its pretraining data and safety…
Short excerpt from the collected text, not the full source. Use the source link to read it in context.
The factual summary is a OneBench synthesis, not a quotation or independent verification. Collection time is not publication time. Open the source for its full context; related reporting can share the same underlying announcement.
OneBench interpretation
Institutional assessment
So what
Red-teaming a single model checkpoint creates false confidence because random seed variations alter specific vulnerability profiles.
Do what
Review safety validation protocols with the team responsible for model risk management to ensure red-teaming covers multi-seed evaluations.