XNews

AI model attribution and security incidents

Model attribution and Hugging Face breach reported. - Aggregator OpenRouter hides backend models. - Tomáš Bruckner proposed fingerprint test. - Method: ask model for a random number. - Models fail to produce true randomness. - GPT-4o favors 42, also 37 and 57. - Claude Sonnet 5 predominantly says 47. - Qwen3-Max answered 42 in 30 samples. - Hugging Face internal network was breached. - Intrusion spanned a weekend with 17,000+ logs. - Their agent was blocked by a paid model. - Resolution used self-hosted GLM 5.2 model.

Key insights

Random-number fingerprinting method for model attribution

Tomáš Bruckner proposed a fingerprinting method that asks a model to produce a random number and identifies the model by its non-random answer patterns. The method relies on the observation that large models do not produce true randomness and instead favor specific numbers. Bruckner’s technique treats those preferred numbers as a kind of model fingerprint that reveals which model answered, even when the caller only selected a name at an aggregator entry point.

Observed model-specific number preferences documented

Empirical examples cited show model-specific preferences: GPT-4o most often answers 42 (also 37 and 57), Claude Sonnet 5 predominantly answers 47, and Qwen3-Max returned 42 in all 30 sampled queries. These consistent numeric preferences are presented as stable signals usable for model identification in aggregator setups.

Hugging Face internal network intrusion and response

Hugging Face reported an internal network intrusion that spanned a weekend and produced more than 17,000 action logs. During investigation, their internal agent was treated as an attacker by a paid model and was blocked. The team ultimately resolved the incident using a self-hosted environment running the open-source GLM 5.2 model.

Source posts