Metraipolis › archive › dc_6d2a2a65041a

Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injection — arXiv:2609.15989

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.15989
License (as archived)
CC0-1.0
Archived
2026-10-15T22:48:11.164Z
SHA-256 of the archived text
6d8953eaf58c0ecbdebc511f183ede6165e4b383f832dfd4827e74d0d8186a13
Stable URL of this page
https://metraipolis.com/documento/dc_6d2a2a65041a

How to cite

Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injection — arXiv:2609.15989 — arXiv (en), https://arxiv.org/abs/2609.15989 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_6d2a2a65041a (Metraipolis archive dc_6d2a2a65041a, sha256 6d8953eaf58c0ecbdebc511f183ede6165e4b383f832dfd4827e74d0d8186a13)
BibTeX
@misc{dc6d2a2a65041a,
  title = {Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injection — arXiv:2609.15989},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.15989},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_6d2a2a65041a; sha256 6d8953eaf58c0ecbdebc511f183ede6165e4b383f832dfd4827e74d0d8186a13},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.15989. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Chain-of-thought (CoT) monitoring is a safety strategy where the reasoning of a large language model "actor" is inspected by a "monitor" (often another language model) for signs of unsafe planning, deception, or misalignment. We find that planting harmful but benign-sounding reasoning in the actor's context can steer it to perform adversarial actions while evading monitors, an attack we term "plan injection". We initially discover this attack in the multiple-choice question-answering monitorability setting proposed by Lanham et al. (2023), using the investigator-agent elicitation framework of Li et al. (2025). We generalize the attack and show that the discovered behavior scales to harder tasks (achieving 25-33% monitor evasion rates across different monitorability benchmarks) and larger models such as DeepSeek-R1. Across the settings we study, actor models not only follow injected plans but also paraphrase them as their own reasoning, without explicit attribution to the injections. Finally, we find cases where extra monitor resources cause harm - giving the monitor access to the injected plan drops detection by as much as 50% in the Bio-Math task and in a case study on monitor reasoning budget, we find transcripts where additional thinking tokens are spent rationalizing the injected plan rather than flagging it.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-15 in cs.AI).
Categorie arXiv: Artificial Intelligence (cs.AI).
Pagina del lavoro: https://arxiv.org/abs/2609.15989 — 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 2026-10-15T22:29:03.383Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-09-15?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).