New March 2026 Paper Exposes Fraudulent Third‑Party APIs for Large Language Models

A newly posted arXiv paper (arXiv:2603.01919v1) investigates the rapidly growing ecosystem of “shadow APIs” that promise cheaper or unrestricted access to cutting‑edge large language models such as GPT‑5 and Gemini‑2.5, bypassing pricing, payment, or regional restrictions.

The authors compiled a list of 17 shadow‑API services that have been referenced in 187 academic publications. They evaluated each service along three dimensions—Utility, Safety, and Model Verification. One of the most widely used projects on GitHub has nearly 60 000 stars and over 5 900 citations, underscoring the prevalence of these services in research.

Experimental results reveal a substantial performance discrepancy between official APIs and their shadow counterparts, with a maximum deviation of 47.21%. For instance, on the high‑risk medical benchmark MedQA, the official Gemini‑2.5‑flash model achieves 83.82% accuracy, whereas the tested shadow API’s accuracy collapses to roughly 37.00%.

The paper includes detailed tables and charts (see the figure below) that document these gaps; the article’s author chooses not to reproduce the full data but encourages readers to consult the original study.

These findings raise serious concerns about the reliability of downstream applications that depend on shadow APIs and threaten the reproducibility of scientific research that assumes parity with official models. The author also criticizes naïve responses that ignore these risks, noting that the lack of trustworthy third‑party services hampers broader development.