On October 21–22, 2026, Cybernews researchers disclosed that 302.ai had left a Microsoft Azure storage bucket publicly accessible. 302.ai is an AI "aggregator" that resells access to ChatGPT, Claude, Grok, DeepSeek, Gemini and other large language models through one interface. The bucket held 17.6 million files totalling 22.72TB of user data. According to Insurance Business, 302.ai is developed by Univerads Technology Limited, a Hong Kong company founded in 2020. Among the exposed enterprise data were several conversations from a customer-service chatbot run by what Cybernews called "a prominent sex doll manufacturer." The company was not named.

Why It Matters

Doll and companion-hardware makers are adopting LLMs quickly, both for customer support and inside the products. Afterglow has covered [voice-enabled dolls](/stories/dollpimp-evas-doll-voice-legs) and [AI-driven bionic inserts](/stories/irontech-ironai-bionic-vaginax). Many of these companies are small. Instead of contracting directly with OpenAI or Anthropic, they often route traffic through cheap third-party gateways like 302.ai, which adds another company holding their data. For most businesses a leaked support chat is embarrassing. For a sex-doll brand, a support chat can contain a customer's name, address, order details and preferences, and that kind of exposure is a ready-made extortion script. This is also the third kind of intimate-tech data failure we've tracked this year. [Lovense's account-takeover flaws](/stories/lovense-20m-user-security-fallout) were in the device platform itself. The [AI girlfriend apps leaking user data](/stories/ai-girlfriend-apps-security) were companion apps. Here the weak point is the AI vendor's vendor. Brands should ask three things: which gateway their chatbot runs through, where its logs are stored, and how long they're kept. Under GDPR, sexual-life data is a special category. A misconfigured bucket at a Hong Kong reseller won't shield a European brand from liability if its customers' details turn up in a dump.

The dataset was mostly generated media: nearly 9TB of images, 7.7TB of video and 5.6TB of audio. The more sensitive part was more than 144,000 chat logs stored as JSON. Each log held the user's prompt, the model's reply, which model was used, token costs, timestamps and other metadata. Some logs contained personal details users had typed in themselves, from birth dates and gender to contact details and home addresses. The bucket also held more than 26,000 .zip archives of conversation data and documents users had uploaded for analysis.

Insurance Business, citing the Cybernews timeline, says the leak was found on August 2, reported to the company on August 11, and the bucket was not seen closed until September 8. The data is now offline. Cybernews says it has no evidence that malicious actors accessed it, but added that "if our team was able to find it, less impartial individuals may have, too." The researchers warned that leaked chats are useful raw material for personalised social-engineering attacks. They also argued that responsibility for securing AI conversations is shared by the businesses that deploy agents, not only the model developers.

Sources

Update — 2026-10-23

Initial entry — story first created.