Work

Fluency as Attack Surface - language-framing and reasoning-trace bleed in frontier LLMs

Year: 2026

Type: article

Abstract: An empirical study of what twelve current LLMs do when prompted innine target languages (real and constructed) plus an English baseline.960 cells of in-language responses, 108 self-translations, 108 externaltranslations, four pipeline phases, 487 cells flagged-and-recoveredbetween v1 and v4. Findings on reasoning-trace leakage, safety-classifiermis-triggers, training-data gaps, cross-language commitment-directionsteering, and constraint-prompt limits, with raw-data appendices.

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.20273325

Author: Ptaszyński, Bartosz

Language: English

Topic: Natural Language Processing Techniques

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7161583336

FWCI:

Cites:

Cited by:

Related to: