Contagious Fingerprints: When LLMs Inherit Another Model's Identity
Organizations: Nanyang Technological University
Abstract
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 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 Lineup, which samples references from the claimed model on the same history and tests the audited reply by its rank among them. 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.
Figures & tables
| passive threshold | Lineup | ||||
|---|---|---|---|---|---|
| Score | Sessions | ||||
| Neural | substitutes | 0.55 | 0.77 | 0.25 | 0.82 |
| honest | 0.07 | 0.09 | 0.07 | 0.05 | |
| N-gram | substitutes | 0.38 | 0.55 | 0.20 | 0.75 |
| honest | 0.06 | 0.07 | 0.06 | 0.06 | |
| Method | Camouflaged substitute | Honest model on long history | Requires | Visible to relay |
|---|---|---|---|---|
| Passive classifier, fixed threshold | missed (0.54–0.72, ) | ok | text only | no |
| + handoff training | seen substitutes only | ok | handoff data of the substitute | no |
| contemporaneous calibration | 0.77 (closed, ) | ok (0.09) | honest sessions of same config. | no |
| Random-number probes, fresh reference | detected (1.00) | flagged (0.24–1.00 in 5/9 models) | fresh samples | yes (0.98) |
| Random-number probes, official matched reference | detected (0.89–1.00) | ok | official calls | yes |
| MET-style MMD, fresh reference | – | flagged (0.55–1.00, 9/9 models) | fresh samples | yes |
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| ID | Contrast | Neural | N-gram |
|---|---|---|---|
| H1a | Markdown removed, downgrades | 4.62 [3.49, 5.73] | 1.21 [0.87, 1.59] |
| H1b | downgrades local | 3.56 [2.28, 4.81] | 0.50 [ 0.03, 1.05] † |
| H2 | Third-party restatement | 2.73 / 4.33 | 1.51 / 1.60 |
| H3a | Own-style restatement, local | 3.61 [3.11, 4.09] | 1.81 [1.43, 2.15] |
| H3b | Third-style restatement | 1.96 [1.57, 2.34] | 1.30 [1.00, 1.59] |
| H4 | Recency share, local / API | 0.32 / 0.41 † | 0.21 / 0.37 † |
| N-gram | Neural | |||||
|---|---|---|---|---|---|---|
| Handoff | own | after | +train | own | after | +train |
| Sonnet Haiku | 0.51 | 0.24 | 0.32 | 0.87 | 0.54 | 0.62 |
| Sonnet DeepSeek | 0.87 | 0.45 | 0.78 | 0.97 | 0.72 | 0.83 |
| Sol DeepSeek | 0.83 | 0.43 | 0.72 | 0.98 | 0.92 | 0.96 |
| Sol Luna | 0.51 | 0.38 | 0.52 | 0.97 | 0.94 | 0.96 |
| Sonnet MiniMax ∗ | 0.94 | 0.60 | 0.57 | 0.97 | 0.55 | 0.55 |
| Setting | Sessions | ||||
|---|---|---|---|---|---|
| 16 turns, honest | 80 | .06 | .01 | .00 | .03 |
| 16 turns, handoff | 160 | .20 | .87 | 1.00 | 1.00 |
| 16 turns, handoff | 160 | .22 | .89 | .99 | – |
| Agent prompt, honest | 80 | .05 | .03 | – | – |
| Agent prompt, handoff | 160 | .19 | .81 | – | – |
| Real users, honest | 120 | .06 | .07 | – | – |
| Lineup | passive, | |
|---|---|---|
| Sonnet, honest | 0.06 | 0.06 |
| Sonnet Haiku | 0.78 | 0.70 |
| Sonnet DeepSeek | 0.62 | 0.62 |
| Sonnet MiniMax | 0.93 | 0.95 |
| Sol, honest | 0.05 | 0.11 |
| Sol Luna | 0.97 | 0.93 |
| closed | local | |||
|---|---|---|---|---|
| Test | ||||
| Lineup -MMD, RBF on neural embeddings | .03 / .86 | .05 / 1.00 | .05 / .91 | .05 / 1.00 |
| Lineup -MMD, character Hamming | .03 / .30 | .08 / .58 | .04 / .33 | .06 / .67 |
| Lineup , neural score | .04 / .42 | .05 / .82 | – | .05 / .92 |
| closed | local | ||
|---|---|---|---|
| Score | honest | substitute | substitute |
| N-gram | .07 / .07 | .76 / .81 | .89 / .97 |
| Neural | .04 / .06 | .79 / .89 | .92 / .98 |
| Stylometric | .06 / .05 | .61 / .67 | .78 / .84 |
| N-gram + neural | .06 / .05 | .85 / .92 | .93 / .99 |
| Likelihood | – | – | .94 / 1.00 |
| Model | Vendor | Weights | $/M in / out | Class |
|---|---|---|---|---|
| Claude Sonnet 5.5 † | Anthropic | closed | 2.00 / 10.00 | yes |
| GPT-6.1 Sol † | OpenAI | closed | 2.00 / 10.00 | yes |
| Claude Haiku 5.5 | Anthropic | closed | 0.10 / 0.50 | yes |
| GPT-6 Luna | OpenAI | closed | 0.10 / 0.50 | yes |
| DeepSeek V4.1 Flash | DeepSeek | API | 0.15 / 0.60 | yes |
| MiniMax M3 | MiniMax | API | 0.30 / 1.20 | no |
Language Model Fingerprinting
Language model fingerprinting identifies and verifies models or their lineage from distinctive outputs, internal representations, or embedded signals.
See all 87 papers