CANLI
xAI, Imagine API’yi 2.0’a Yükseltmeye Hazırlanıyor: Görüntü ve Video Tek…·Microsoft MAI-Cyber-1-Flash’ı Duyurdu·Moonshot AI, Kimi K3 Model Ağırlıklarını ve Teknik Raporunu Açık…
14 Nov 2026 · 00:10 GMT+3
Ai Haber – Türkiyenin Yapay Zeka Haber Portalı
ARAşTıRMA · DOğAL DIL İşLEME arXiv:2611.17280 12 Kas 2026 · v1

Synthesis Smuggling: Maliciously Hiding Inferred Personal Attributes in LLM Outputs

Gautam Galada, Supreeth Shastri, Sagnik Ray Choudhury

YAYIN:12 Kas 2026 ALAN:cs.CL

Özet

We propose Synthesis Smuggling, a novel threat model in which an adversarial LLM infers sensitive information about a user from routine assistance-seeking interactions and steganographically encodes that into innocuous-looking model-generated content. The user may then unknowingly transmit that information by sharing the generated artifact, say as an email, social-media post, or message, through ordinary communication channels. An adversary monitoring these channels can discover this information. We demonstrate this threat by training two open-weight models to infer gender and age-group labels from conversational cues and encode their predictions using invisible characters or word choices. These results show that user-shared model outputs can become a covert channel for personal inferences, even when the model itself has no direct means of contacting an adversary.

Özetle: We propose Synthesis Smuggling, a novel threat model in which an adversarial LLM infers sensitive information about a user from routine assistance-seeking interactions and steganographically encodes that into innocuous-…

Özet

We propose Synthesis Smuggling, a novel threat model in which an adversarial LLM infers sensitive information about a user from routine assistance-seeking interactions and steganographically encodes that into innocuous-looking model-generated content. The user may then unknowingly transmit that information by sharing the generated artifact, say as an email, social-media post, or message, through ordinary communication channels. An adversary monitoring these channels can discover this information. We demonstrate this threat by training two open-weight models to infer gender and age-group labels from conversational cues and encode their predictions using invisible characters or word choices. These results show that user-shared model outputs can become a covert channel for personal inferences, even when the model itself has no direct means of contacting an adversary.

Orijinal Özet (İngilizce)

We propose Synthesis Smuggling, a novel threat model in which an adversarial LLM infers sensitive information about a user from routine assistance-seeking interactions and steganographically encodes that into innocuous-looking model-generated content. The user may then unknowingly transmit that information by sharing the generated artifact, say as an email, social-media post, or message, through ordinary communication channels. An adversary monitoring these channels can discover this information. We demonstrate this threat by training two open-weight models to infer gender and age-group labels from conversational cues and encode their predictions using invisible characters or word choices. These results show that user-shared model outputs can become a covert channel for personal inferences, even when the model itself has no direct means of contacting an adversary.

Kaynak: arXiv:2611.17280 · PDF

BibTeX

@article{galada2026synthesis,
  title   = {Synthesis Smuggling: Maliciously Hiding Inferred Personal Attributes in LLM Outputs},
  author  = {Gautam Galada and Supreeth Shastri and Sagnik Ray Choudhury},
  journal = {arXiv preprint arXiv:2611.17280},
  year    = {2026},
  url     = {https://arxiv.org/abs/2611.17280}
}

Tartışma

Bu habere emoji ile tepki ver

Hizli:

Henüz yorum yok. İlk yorumu siz yapın!

Yapıcı ve saygılı yorumlar bekliyoruz. Topluluk kuralları