Evidence · DiggingBeagle record
STEALTH demonstrates model-output steganography across GPU and model configurations (research paper)
Keyed sampling and error-correcting text coding can transmit a small covert payload through generated language despite some model and hardware mismatches.
- Published
- Nov 12, 2026
- Source role
- primary disclosure
Evidence record
Keyed sampling and error-correcting text coding can transmit a small covert payload through generated language despite some model and hardware mismatches.
Claim-level citations (4)
- supportsSTEALTH demonstrates model-output steganography across GPU and model configurations: The STEALTH experiments vary the sender and receiver model families, precision modes and GPU implementations including an RTX A6000 and an H100. These controlled changes probe decoder portability across model-serving configurations, not extraction from another tenant's GPU or demonstrated delivery of stolen real-world secrets.
Experimental setup, hardware and precision comparison of sender and receiver configurations
- supportsSTEALTH demonstrates model-output steganography across GPU and model configurations: The authors reach full message recovery for some cross-family pairs only when sufficiently long complete output streams are available; at the sender's initial stopping point, selected pairs recover only 12.5 to 19 percent.
Evaluation, cross-model and early-stopping recovery tables
- supportsSTEALTH demonstrates model-output steganography across GPU and model configurations: The sender uses keyed exponential-race sampling and rateless coding, while the receiver performs beam decoding across noisy token distributions and model-specific tokenization differences.
Sampling and rateless coding methodology
- supportsSTEALTH demonstrates model-output steganography across GPU and model configurations: The STEALTH evaluation distinguishes full-stream recovery from the sender's initial stopping length: some cross-family pairs recover all messages only with additional tokens, while selected pairs at early stopping recover 12.5 to 19 percent.
Section VII, cross-family decoding, stream length and stopping-point recovery results