Case · DiggingBeagle record

TrojanStego trained a language model to leak secrets inside natural-looking text

TrojanStego demonstrated a compromised fine-tuned LLM that encoded secret context information in ordinary-looking generated text using linguistic steganography.

LLM steganographic communication Human-imperceptible machine-readable channels

Controlled compromised-model threat model. It does not show that ordinary deployed LLMs spontaneously exfiltrate secrets this way.

Case kind: vulnerability
Claims: 2

Reconstruction

Timeline

2025-11
Step

Claims & evidence

reported findingsupported

The paper reports reliable transmission of 32-bit secrets, with 87% accuracy on held-out prompts and above 97% using majority voting across three generations in its experiments.

supports
TrojanStego: Your Language Model Can Secretly Be A Steganographic Privacy Leaking Agent
Locator: ACL Anthology abstract, reported accuracy

reported findingsupported

TrojanStego fine-tuned an LLM to encode sensitive context information into natural-looking outputs through a vocabulary-partitioning steganographic scheme.

supports
TrojanStego: Your Language Model Can Secretly Be A Steganographic Privacy Leaking Agent
Locator: ACL Anthology abstract

Implications

Surface-fluent text is not proof that the output carries only its visible semantic payload.

Controls & mitigations

Model provenance and weight integrity controls
Output transformation or randomized paraphrasing as one layer, not a complete defense
Statistical detection of abnormal token-choice patterns
Restrict sensitive context exposure to untrusted models

What remains unknown

Robustness under active rewriting, production moderation and heterogeneous downstream decoders remains separate from the controlled experimental result.

Cite this record

DiggingBeagle. “TrojanStego trained a language model to leak secrets inside natural-looking text.” https://diggingbeagle.com/cases/trojanstego-trained-a-language-model-to-leak-secrets-inside-natural-looking-text/

Citation guidance