An AI aggregator called 302.ai left 22.72TB of user and corporate data in a publicly reachable cloud storage bucket, according to researchers at Cybernews.
An aggregator is a service that puts several AI models behind one interface, so a user can switch from ChatGPT to Claude without leaving the page. The researchers said 302.ai, which Hong Kong-based Univerads Technology Limited launched six years ago, kept a Microsoft Azure bucket (a cloud storage container) with no password and no encryption. It held 17.6 million files. Images made up almost 9TB, video 7.7TB and audio 5.6TB. The remaining 500GB or so was more than 144,000 chat logs saved as .json files.
Those logs recorded user prompts, the AI replies, which model answered, the token cost and timestamps. Some users had also typed in names, birth dates, genders, postal addresses and other contact details. The researchers also found logs from a customer support chatbot run by a sex doll manufacturer, which shows businesses were using the service for real work.
Cybernews found the bucket in early August 2026 and said it was locked down about a month later. Univerads did not answer the researchers' messages. It remains unclear who maintains the database, when collection began, whether anyone else found the files first, or whether anything was copied.
The cause here is an ordinary misconfiguration, not anything peculiar to AI. What AI adds is the content: people treat a chat box like a private notepad, and an aggregator stores those notes for every model it routes to. That puts a middleman's storage hygiene between a user and their own prompts.
The data is also useful for fraud. Knowing which company a person works for and what they were discussing makes a phishing email far more convincing, and such emails often lead to business email compromise or ransomware. Large cloud exposures of this kind are very common, the source noted, and one in Brazil affected entire populations. For 302.ai's users, a month-long window with no public answer on access means the safest assumption is that anything they typed there may have been seen.