302.ai leak exposed 17.6 million files and sex doll support chats

A security lapse at AI aggregator 302.ai exposed a publicly accessible Azure storage bucket containing 17.6 million files and 22.72TB of user data. The leak included chat logs, generated images, videos, audio files, documents, and archived conversation files. After researchers contacted the company, the exposed data was taken offline.

What the leak exposed

The bucket held 17.6 million files totaling 22.72TB. Images made up nearly 9TB, video about 7.7TB, and audio files about 5.6TB. The report said it was highly likely the data was left exposed accidentally, though that was not confirmed.

Chat logs and personal data

More than 144,000 chat logs were stored in .json files. They exposed user prompts, assistant replies, the model used, token costs, timestamps, and other metadata. Some logs contained personally identifiable information that users likely provided themselves, including birth dates, genders, contact information, and home addresses.

Enterprise files and conversation archives

The leak also included enterprise client information. Researchers found multiple chat logs from a customer support chatbot owned by a prominent sex doll manufacturer. Other exposed material included numerous documents, some discussing LLM development research and others resembling private documents users likely wanted to analyze. The bucket also contained more than 26,000 .zip files with conversation data, including conversation summaries, user input and output by the model, and files included in conversations.

Timeline and company background

The leak was discovered on August 2, 2026, disclosed on August 11, 2026, and observed closed on September 8, 2026. 302.ai is developed by Univerads Technology Limited, a Hong Kong-based company founded in 2020. 302.ai lets customers access major large language models such as ChatGPT, Claude, Grok, DeepSeek, and Gemini. The report said there was no evidence that malicious actors accessed or downloaded the data, but warned that leaked chats could help personalized social engineering campaigns.

Sources: Cybernews