Every source cited on this claim, counted once at its highest evidence strength and grouped by the side it supports. Generated from this page's own evidence rows — the same records the verdict is computed from — so the chart and the score cannot disagree. Strength labels follow the scoring methodology.

View figure at source: Stealing Machine Learning Models via Prediction APIs

Tramèr et al. (2016) model-extraction results showing how prediction-query attacks reconstruct machine-learning models, with extraction fidelity or accuracy plotted against the number of API queries a

The clearest landmark visualization of how proprietary model behavior can be replicated remotely without direct access to model-weight files. It directly illustrates a plausible diffusion pathway, while also distinguishing model extraction from confirmed theft of frontier-laboratory weights.

View figure at source: An Empirical Analysis of Memorization in Fine-tuned Autoregressive Language Models

Carlini et al. (2022) memorization experiments showing the relationship between training conditions and verbatim recovery from fine-tuned autoregressive language models, with memorization or extractio

This figure provides the most relevant visual evidence for information leakage from language models: distinctive or repeated training examples can be reproduced from model behavior. It helps readers distinguish data memorization from exfiltration of an entire training corpus or model-weight theft.

View figure at source: The H-Elena Trojan Virus to Infect Model Weights: A Wake-Up Call on the Security Risks of Malicious Fine-Tuning

H-Elena experimental results showing the effect of malicious fine-tuning on model-weight integrity and downstream model behavior, comparing clean and compromised models across attack success, task per

This is a concrete visual demonstration that model weights can be compromised through a supply-chain or fine-tuning pathway. It broadens the debate beyond theft to include unauthorized alteration and diffusion of compromised model artifacts, while not establishing that leading laboratories have experienced such incidents.

All contributions are reviewed for clarity, balance, and evidence. The strongest insights are elevated into the argument graph — with credit to you.

Help improve this analysis →

𝕏 Share Facebook LinkedIn