Case · 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.

Scope

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.

At a glance

Mechanism and trust boundary

  1. 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

  2. 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

  3. 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

Link to claim
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

Link to claim
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

Link to claim
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

Link to claim

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

  1. Nov 15, 2026 · Canonical change recorded · new in release · revision 251