{"slug": "contagious-fingerprints-when-llms-inherit-another-model-s-identity", "title": "Contagious Fingerprints: When LLMs Inherit Another Model's Identity", "summary": "A model that continues a conversation started by another model inherits that model's fingerprint, undermining audits of API relays that quietly swap in cheaper models, according to arXiv paper 2611.18104v1. Detection of a substitute by a text classifier fell from 0.97 to 0.72 when DeepSeek replaced Claude Sonnet and from 0.87 to 0.54 for a same-vendor downgrade to Haiku, and retraining on handoff data repaired only substitutes seen in training. The authors propose Lineup, which samples references from the claimed model on the same conversation history and ranks the audited reply among them; with contemporaneous references on two closed models it reported a false-positive rate of 0.056 against a preregistered bound of 0.10 and detected 0.81 of substitute sessions across six pairs after four audited turns.", "body_md": "arXiv:2611.18104v1 Announce Type: new \nAbstract: The fingerprint by which a classifier recognizes the LLM behind a text is usually treated as a fixed property of the model. We show that it is contagious: a model that continues a conversation begun by another model inherits that model's fingerprint. This undermines audits of API relays, which can serve the first turns of a conversation with the model a customer pays for and silently hand the rest to a cheaper one. After such a handoff, detection of the substitute by a text classifier drops from 0.97 to 0.72 (DeepSeek replacing Claude Sonnet) and from 0.87 to 0.54 for a same-vendor downgrade (Haiku), and retraining on handoff data repairs only substitutes seen in training. Preregistered experiments trace the contagion to imitation of the history's style. History also moves a model away from its own fingerprint: active probes are not camouflaged, but when calibrated on fresh conversations they flag \\emph{honest} models on long contexts at rates up to 1.00, and a standard MMD equality test rejects all nine models we test. Both failures come from comparing a history-dependent reply with a history-free reference. We propose \\textsc{Lineup}, which samples references from the claimed model \\emph{on the same history} and tests the audited reply by its rank among them. \\textsc{Lineup} controls false positives by exchangeability and needs no fitted threshold or log-probabilities. With contemporaneous references on two closed models, its false-positive rate is 0.056 (preregistered bound 0.10), and it detects 0.81 of substitute sessions across six pairs after four audited turns.", "url": "https://wpnews.pro/news/contagious-fingerprints-when-llms-inherit-another-model-s-identity", "canonical_source": "https://www.machinebrief.com/news/contagious-fingerprints-when-llms-inherit-another-models-ide-pwa6", "published_at": "2026-11-16 05:00:00+00:00", "updated_at": "2026-11-16 06:22:08.193660+00:00", "lang": "en", "topics": ["ai-safety", "large-language-models", "ai-research", "artificial-intelligence"], "entities": ["DeepSeek", "Claude Sonnet", "Claude Haiku", "Lineup", "arXiv"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/contagious-fingerprints-when-llms-inherit-another-model-s-identity", "markdown": "https://wpnews.pro/news/contagious-fingerprints-when-llms-inherit-another-model-s-identity.md", "text": "https://wpnews.pro/news/contagious-fingerprints-when-llms-inherit-another-model-s-identity.txt", "jsonld": "https://wpnews.pro/news/contagious-fingerprints-when-llms-inherit-another-model-s-identity.jsonld"}}