Security firm Claroc discovered that the AI application builder from SpaceXSI failed to protect the privacy of the vast majority of its users’ hosted projects. An analysis of 37,571 live applications revealed that 35,578 were open to the public without requiring a sign-in, including 3,784 worker medical exams and a roster for a Navy contractor. The breach also exposed 24 internal apps created by staff at X and SpaceXSI that contained advertising proposals and a territory map of 594 accounts.
The exposure occurred because the platform used random strings in web addresses that passive DNS services log, unlike Gemini, ChatGPT, and Claude, which place secrets in the URL path to avoid such logging. While a representative for Cursor, a company acquired by SpaceXSI, stated that apps are private by default, Claroc’s tests showed that paid accounts had a setting for “Anyone with the link” preselected. SpaceXSI removed its staff applications from the web after the findings were reported by Ars Technica.
