Metraipolis › archive › dc_ccadef46585b

Erase the Injected Region: Multimodal Prompt-Injection Defense with On-Policy Self-Distillation — arXiv:2610.41307

Source
arXiv (en)
Original
https://arxiv.org/abs/2610.41307
License (as archived)
CC0-1.0
Archived
2026-11-03T05:33:07.378Z
SHA-256 of the archived text
dce8aa0efc382ade5984f608b442dd08efe46bdab0b41d1ffec8fb7cec2c3dbb
Stable URL of this page
https://metraipolis.com/documento/dc_ccadef46585b

How to cite

Erase the Injected Region: Multimodal Prompt-Injection Defense with On-Policy Self-Distillation — arXiv:2610.41307 — arXiv (en), https://arxiv.org/abs/2610.41307 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_ccadef46585b (Metraipolis archive dc_ccadef46585b, sha256 dce8aa0efc382ade5984f608b442dd08efe46bdab0b41d1ffec8fb7cec2c3dbb)
BibTeX
@misc{dcccadef46585b,
  title = {Erase the Injected Region: Multimodal Prompt-Injection Defense with On-Policy Self-Distillation — arXiv:2610.41307},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2610.41307},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_ccadef46585b; sha256 dce8aa0efc382ade5984f608b442dd08efe46bdab0b41d1ffec8fb7cec2c3dbb},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.41307. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Multimodal indirect prompt injection embeds malicious instructions in external media that also contains information needed for the user's task. Existing defenses can block the attack yet impair the model's use of that information, leaving a gap between attack defense and task preservation. We propose ERASE, an on-policy self-distillation method that constructs its reference by erasing the injected region in place. Modality-appropriate blank content replaces the injection while the surrounding media is preserved. A student reading attacked media learns from a frozen base model reading the erased view through full-distribution token-level feedback on its own rollouts. Across image, audio, and video on two model sizes, ERASE records near-zero observed attack success on the evaluated benchmark while retaining task utility close to the undefended clean-input level. On 3B audio, ERASE reaches 60.1% attacked-input accuracy compared with ARGUS's 33.1%, within 0.5 points of the undefended clean-input level. Experiments show that ERASE reduces injection dependence from 6.26 to 0.0095 on 3B audio while staying close to the base model on the erased input. Out-of-domain accuracy also stays near the base level. Our code will be made publicly available upon acceptance.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.AI, nuovi lavori del giorno).
Categorie arXiv: Cryptography and Security (cs.CR) ; Artificial Intelligence (cs.AI).
Pagina del lavoro: https://arxiv.org/abs/2610.41307 — 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 2026-11-03T05:18:52.206Z dalla pagina https://arxiv.org/list/cs.AI/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).