Keyed sampling and error-correcting text coding can transmit a small covert payload through generated language despite some model and hardware mismatches. 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. The authors…
Inspect the ClaimsCase · DiggingBeagle record
STEALTH demonstrates model-output steganography across GPU and model configurations
Keyed sampling and error-correcting text coding can transmit a small covert payload through generated language despite some model and hardware mismatches.
Controlled news-comment generation across selected Gemma, Llama and Qwen models and GPU/precision configurations; no observed illicit exfiltration.
Assessment profile
Assessment method- Setting
- evaluation
- Exploitation
- demonstrated
- Evidence
- Grade C
- Remediation
- not applicable
- AI role
- WITH AI
Basis and provenance
Some sufficiently long streams recover all tested payloads, but early stopping sharply reduces success and no hostile data exfiltration has been established. Classified as evaluation with exploit status demonstrated. Evidence grade C reflects the original paper or maintainer advisory and its stated limitations, not an independent audit of all deployments. Comparable realized harm is not established, so impact stays unrated; remediation status not_applicable is limited to the cited experiment or affected software version. New source-specific boundary: Model-output watermark recovery and covert-message recovery should be evaluated at the actual delivered stream length and receiver configuration; hardware variability alone does not establish operational exfiltration.
Assessed Nov 15, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-R245-STEALTH-LLM-COVERT-OUTPUT-01 · CLM-R245-STEALTH-LLM-COVERT-OUTPUT-02 · CLM-R246-STEALTH-LLM-COVERT-OUTPUT-BOUNDARY · CLM-ENR250-17 · STEALTH demonstrates model-output steganography across GPU and model configurations (research paper)
At a glance
The full-stream success headline depends on extra output length and a cooperating receiver; the study does not prove survival against arbitrary paraphrasing, translation or content rewrites. No evidence here establishes the number of production victims or a broader failure rate outside the cited evaluation.
Limits and uncertaintyFull account
Keyed sampling and error-correcting text coding can transmit a small covert payload through generated language despite some model and hardware mismatches.
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. 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.
The full-stream success headline depends on extra output length and a cooperating receiver; the study does not prove survival against arbitrary paraphrasing, translation or content rewrites.
A cooperating sender can use a secret key to choose natural-language tokens whose stochastic distribution carries a coded payload. STEALTH combines exponential-race sampling with rateless spinal coding and a beam-search receiver, allowing the receiver to tolerate differences between GPU computations and tokenizers. Unlike a normal prompt injection, the message may be hidden in text that appears innocuous to a reader, although a recipient must have the decoding key and sufficient output length.
Hardware mismatch is not a complete defense, but neither is the claimed full-stream decoding rate evidence of reliable short-message extraction. Several cross-model comparisons need hundreds of extra tokens beyond the original stop length before all messages can be reconstructed. The experiments concern controlled news-comment generation, not observed stolen secrets, robust survival through translation or an independently audited production covert channel. A security evaluation must report bits per token, actual delivered length, decoding time and post-processing separately.
Mechanism and trust boundary
- 01
Entry or starting condition
A cooperating model-output sender seeks to convey hidden bits through an otherwise ordinary generated comment.
Boundary: attacker or research precondition
- 02
Crossing the relevant boundary
Keyed sampling and rateless error-correction encode bits into token choices, and a key-holding receiver decodes the stream under hardware or tokenizer mismatch.
Boundary: mechanism and authorization boundary
- 03
Observed or reported result
Some sufficiently long streams recover all tested payloads, but early stopping sharply reduces success and no hostile data exfiltration has been established.
Boundary: measured outcome and evidentiary limit
Claims & evidence
CLM-ENR250-17The 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.supported
Basis: reported finding
- supportsSTEALTH demonstrates model-output steganography across GPU and model configurations (research paper)primary disclosure
Experimental setup, hardware and precision comparison of sender and receiver configurations
CLM-R245-STEALTH-LLM-COVERT-OUTPUT-01The 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.supported
Basis: reported finding
- supportsSTEALTH demonstrates model-output steganography across GPU and model configurations (research paper)primary disclosure
Evaluation, cross-model and early-stopping recovery tables
CLM-R245-STEALTH-LLM-COVERT-OUTPUT-02The sender uses keyed exponential-race sampling and rateless coding, while the receiver performs beam decoding across noisy token distributions and model-specific tokenization differences.supported
Basis: reported finding
- supportsSTEALTH demonstrates model-output steganography across GPU and model configurations (research paper)primary disclosure
Sampling and rateless coding methodology
CLM-R246-STEALTH-LLM-COVERT-OUTPUT-BOUNDARYThe 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.supported
Basis: reported finding
- supportsSTEALTH demonstrates model-output steganography across GPU and model configurations (research paper)primary disclosure
Section VII, cross-family decoding, stream length and stopping-point recovery results
Unknowns and contradictions
- The full-stream success headline depends on extra output length and a cooperating receiver; the study does not prove survival against arbitrary paraphrasing, translation or content rewrites.
- No evidence here establishes the number of production victims or a broader failure rate outside the cited evaluation.
Sources and citation
Material revision history
- Nov 15, 2026 · Canonical change recorded · new in release · revision 251