Verifying API Authenticity via LLM Random Number Bias

The arXiv paper "One Token Is Enough"(2607.10252) proposes a verification method that leverages the characteristic that LLMs fail to generate random numbers uniformly, using it as a model fingerprint.

While a uniform distribution should have an entropy of 6.64bit, the measured median is only about 1.0bit. This bias is stable per model and shows clear differences between models, allowing it to function as an identifier.

Verification method:

• Send batches of word-level questions such as numbers, colors, and coin flips at temperature 1.0
• Build an output distribution and compare it with a trusted reference using JS divergence
• Under the full setup, EER is 7.3% and AUC is 0.971

The paper also presents a case where an endpoint labeled as an exclusive flagship model on OpenRouter was actually serving Qwen3-235B (fingerprint distance ~0.141).

However, the paper explicitly states that deviation does not necessarily mean fraud. Quantization, hidden system prompts, routing, and rolling upgrades can also widen the distance, so it should be interpreted as a tool that answers "has the behavior changed?"

A free tool has also been released that can be run directly in the browser while keeping the API key local.