Reinforcement learning and prompting experiments show LLMs can transmit multi-bit payloads covertly in text, with undetected accuracy up to about 66% under a monitor that lacks shared secret context.
’0’ if the word was not replaced
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
The Steganographic Potentials of Language Models
Reinforcement learning and prompting experiments show LLMs can transmit multi-bit payloads covertly in text, with undetected accuracy up to about 66% under a monitor that lacks shared secret context.