Lookalike Letter Attack on AI Models Reveals 'Denial of Spend' Vulnerability

A researcher conducted an experiment where they inserted lookalike letters into a legal contract and tested it against seven different GPT and Claude AI models. While none of the models were deceived by the lookalike characters, the attack caused them to consume up to 5.7 times more tokens to process the document, resulting in costs increasing by up to 3.9 times per question. This demonstrates a 'Denial of Spend' attack vector against AI language models.