Inference-output verification result (Rinberg et al., preprint): exfiltratable information and attacker slowdown on Qwen3-30B-A3B

Name: Inference-output verification result (Rinberg et al., preprint): exfiltratable information and attacker slowdown on Qwen3-30B-A3B
License: https://aidatacenterguide.com/terms#intellectual-property

<0.5% exfiltratable; ≥200× slowdownobserved

Value kind	observed — Reported measurements, counts, and specifications keep the precision and scope stated by their source; an exact specification is not treated as a range.
Scope	Preprint result on one 30B open-weight mixture-of-experts model (the paper tests 3B to 30B models); not a deployed product. The paper's 1 TB, 1000 Gb-per-day example mixes bits and bytes and is not used here.
As of	2026-03-12
Source	Rinberg et al., “Verifying LLM Inference to Detect Model Weight Exfiltration,” arXiv:2511.02620v3, 12 March 2026 (preprint) · Abstract and Contributions: Qwen3-30B-A3B result
Review	checking…review by 2027-03-31 · standard cadence
Recorded changes	last 2026-10-16
Claim id	inference-output-verification-exfiltrable-information-slowdown

Where the guide uses it

11.8 Model & Weight Protection (At-Rest, In-Transit, In-Use)

← Full numbers register — every date-stamped figure in the guide, with revision history.