Metraipolis › archive › dc_f50f87366a42

Visible Reasoning and Indirect Prompt-Injection Monitorability Across English, Tamil, and Tanglish — arXiv:2608.15392

Source
arXiv (en)
Original
https://arxiv.org/abs/2608.15392
License (as archived)
CC0-1.0
Archived
2026-10-18T17:19:43.097Z
SHA-256 of the archived text
38ee992638ba68c83d5abab1bf45c64532056bdf4ef12fe31d2ae1e233ff6a0b
Stable URL of this page
https://metraipolis.com/documento/dc_f50f87366a42

How to cite

Visible Reasoning and Indirect Prompt-Injection Monitorability Across English, Tamil, and Tanglish — arXiv:2608.15392 — arXiv (en), https://arxiv.org/abs/2608.15392 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_f50f87366a42 (Metraipolis archive dc_f50f87366a42, sha256 38ee992638ba68c83d5abab1bf45c64532056bdf4ef12fe31d2ae1e233ff6a0b)
BibTeX
@misc{dcf50f87366a42,
  title = {Visible Reasoning and Indirect Prompt-Injection Monitorability Across English, Tamil, and Tanglish — arXiv:2608.15392},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2608.15392},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_f50f87366a42; sha256 38ee992638ba68c83d5abab1bf45c64532056bdf4ef12fe31d2ae1e233ff6a0b},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.15392. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Chain-of-thought monitoring is a potentially useful safety signal, but its reliability across languages and behavioral settings remains uncertain. In a small case study of eight manually verified synthetic scenarios, one model, one annotator, and one deterministic generation seed, I study API-visible reasoning during indirect prompt injection in Sarvam-105B across English, Tamil, and Tanglish. A four scenario pilot found 5/12 injected attack successes without reasoning and 1/11 with reasoning. A preregistered four-scenario follow-up reversed that direction, finding 2/12 attacks without reasoning and 3/12 with reasoning. With only four scenarios per phase, this design cannot distinguish a real reasoning-mode effect from prompt-specific variation or sampling noise. Across 20 non-empty injected-thinking traces, all 17 benign-correct outputs stated an intent to ignore the injection, while all three attack successes stated an intent to follow it. These descriptive observations provide a reproducible case study of behaviorally informative visible reasoning when it is available; they do not establish that reasoning mode improves safety, that visible reasoning is mechanistically faithful, or that the findings generalize beyond this configuration.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-08-18 in cs.AI).
Categorie arXiv: Artificial Intelligence (cs.AI).
Pagina del lavoro: https://arxiv.org/abs/2608.15392 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-18T16:29:13.797Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-08-18?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).