Definition
Side-Channel Attack is a technique for extracting information from a system through incidental signals produced by its implementation rather than through its intended output. In the context of language models, this means exploiting quirks like tokenizer behavior, chat template formatting, or timing artifacts to infer facts — such as a model’s identity or underlying architecture — even when the interface exposes only text tokens and nothing else. It matters because it shows that restricting an interface to a narrow, seemingly safe channel does not guarantee that deeper implementation details stay hidden.