Salvato in:
Dettagli Bibliografici
Autori principali: Karpov, Artem, Adeleke, Tinuade, Cho, Seong Hah, Perez-Campanero, Natalia
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:https://arxiv.org/abs/2505.03439
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866908351850348544
author Karpov, Artem
Adeleke, Tinuade
Cho, Seong Hah
Perez-Campanero, Natalia
author_facet Karpov, Artem
Adeleke, Tinuade
Cho, Seong Hah
Perez-Campanero, Natalia
contents The potential for large language models (LLMs) to hide messages within plain text (steganography) poses a challenge to detection and thwarting of unaligned AI agents, and undermines faithfulness of LLMs reasoning. We explore the steganographic capabilities of LLMs fine-tuned via reinforcement learning (RL) to: (1) develop covert encoding schemes, (2) engage in steganography when prompted, and (3) utilize steganography in realistic scenarios where hidden reasoning is likely, but not prompted. In these scenarios, we detect the intention of LLMs to hide their reasoning as well as their steganography performance. Our findings in the fine-tuning experiments as well as in behavioral non fine-tuning evaluations reveal that while current models exhibit rudimentary steganographic abilities in terms of security and capacity, explicit algorithmic guidance markedly enhances their capacity for information concealment.
format Preprint
id arxiv_https___arxiv_org_abs_2505_03439
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle The Steganographic Potentials of Language Models
Karpov, Artem
Adeleke, Tinuade
Cho, Seong Hah
Perez-Campanero, Natalia
Artificial Intelligence
Cryptography and Security
Machine Learning
The potential for large language models (LLMs) to hide messages within plain text (steganography) poses a challenge to detection and thwarting of unaligned AI agents, and undermines faithfulness of LLMs reasoning. We explore the steganographic capabilities of LLMs fine-tuned via reinforcement learning (RL) to: (1) develop covert encoding schemes, (2) engage in steganography when prompted, and (3) utilize steganography in realistic scenarios where hidden reasoning is likely, but not prompted. In these scenarios, we detect the intention of LLMs to hide their reasoning as well as their steganography performance. Our findings in the fine-tuning experiments as well as in behavioral non fine-tuning evaluations reveal that while current models exhibit rudimentary steganographic abilities in terms of security and capacity, explicit algorithmic guidance markedly enhances their capacity for information concealment.
title The Steganographic Potentials of Language Models
topic Artificial Intelligence
Cryptography and Security
Machine Learning
url https://arxiv.org/abs/2505.03439