Study finds some third-party AI APIs may serve different models than advertised
A study presented at CCS 2026 audited third-party 'shadow APIs' that offer access to well-known AI models and found performance, safety, and identity differences compared with official APIs. In one test, a service advertised as GPT-5 showed a behavioral fingerprint consistent with GLM-4-9B-chat, though researchers stress this reflects statistical similarity rather than direct server inspection. Researchers have also proposed verification methods, such as analyzing model responses to requests for random numbers and the FLIPS technique presented at ICML 2026, which compare response patterns against known models. These methods reported high accuracy in controlled settings but are not universal guarantees across all services. Experts advise users to verify providers, test output quality, and compare endpoints with trusted references when model identity matters.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in