Research

Paper

TESTING February 18, 2026

IndicJR: A Judge-Free Benchmark of Jailbreak Robustness in South Asian Languages

Authors

Priyaranjan Pattnayak, Sanchari Chowdhuri

Abstract

Safety alignment of large language models (LLMs) is mostly evaluated in English and contract-bound, leaving multilingual vulnerabilities understudied. We introduce \textbf{Indic Jailbreak Robustness (IJR)}, a judge-free benchmark for adversarial safety across 12 Indic and South Asian languages (2.1 Billion speakers), covering 45216 prompts in JSON (contract-bound) and Free (naturalistic) tracks. IJR reveals three patterns. (1) Contracts inflate refusals but do not stop jailbreaks: in JSON, LLaMA and Sarvam exceed 0.92 JSR, and in Free all models reach 1.0 with refusals collapsing. (2) English to Indic attacks transfer strongly, with format wrappers often outperforming instruction wrappers. (3) Orthography matters: romanized or mixed inputs reduce JSR under JSON, with correlations to romanization share and tokenization (approx 0.28 to 0.32) indicating systematic effects. Human audits confirm detector reliability, and lite-to-full comparisons preserve conclusions. IJR offers a reproducible multilingual stress test revealing risks hidden by English-only, contract-focused evaluations, especially for South Asian users who frequently code-switch and romanize.

Metadata

arXiv ID: 2602.16832
Provider: ARXIV
Primary Category: cs.AI
Published: 2026-02-18
Fetched: 2026-02-21 18:51

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2602.16832v1</id>\n    <title>IndicJR: A Judge-Free Benchmark of Jailbreak Robustness in South Asian Languages</title>\n    <updated>2026-02-18T19:54:23Z</updated>\n    <link href='https://arxiv.org/abs/2602.16832v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2602.16832v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Safety alignment of large language models (LLMs) is mostly evaluated in English and contract-bound, leaving multilingual vulnerabilities understudied. We introduce \\textbf{Indic Jailbreak Robustness (IJR)}, a judge-free benchmark for adversarial safety across 12 Indic and South Asian languages (2.1 Billion speakers), covering 45216 prompts in JSON (contract-bound) and Free (naturalistic) tracks.\n  IJR reveals three patterns. (1) Contracts inflate refusals but do not stop jailbreaks: in JSON, LLaMA and Sarvam exceed 0.92 JSR, and in Free all models reach 1.0 with refusals collapsing. (2) English to Indic attacks transfer strongly, with format wrappers often outperforming instruction wrappers. (3) Orthography matters: romanized or mixed inputs reduce JSR under JSON, with correlations to romanization share and tokenization (approx 0.28 to 0.32) indicating systematic effects. Human audits confirm detector reliability, and lite-to-full comparisons preserve conclusions. IJR offers a reproducible multilingual stress test revealing risks hidden by English-only, contract-focused evaluations, especially for South Asian users who frequently code-switch and romanize.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AI'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CL'/>\n    <published>2026-02-18T19:54:23Z</published>\n    <arxiv:comment>Accepted in EACL Industry Track Oral, 2026</arxiv:comment>\n    <arxiv:primary_category term='cs.AI'/>\n    <author>\n      <name>Priyaranjan Pattnayak</name>\n    </author>\n    <author>\n      <name>Sanchari Chowdhuri</name>\n    </author>\n  </entry>"
}