A security lapse at AI platform 302.ai left 17.6 million files totalling 22.72 terabytes publicly accessible, including more than 144,000 AI chat logs containing personal information and conversations linked to a prominent sex doll manufacturer, according to an investigation by Cybernews.
Some chat logs contained dates of birth, genders, contact information and home addresses. Researchers said the data was taken offline after they contacted the company. They found no evidence that malicious actors had accessed or downloaded the files.
The exposed archive also included images, videos, audio files, documents and more than 26,000 ZIP files containing conversation data. Cybernews did not identify the sex doll manufacturer.
How the Data Was Exposed
Researchers discovered 17.6 million files totalling 22.72TB in a publicly accessible Microsoft Azure storage bucket owned by 302.ai. Images accounted for nearly 9TB, while videos made up 7.7TB and audio files a further 5.6TB.
Cybernews said it could not be certain how the data had been left accessible, although its researchers considered accidental exposure likely. The investigation did not establish that criminals had obtained the files.
What the Chat Logs Revealed
Cybernews found more than 144,000 chat logs stored as JSON files. These records included user prompts, assistant responses, the AI model used, timestamps, token costs and other metadata.
Some conversations also contained personally identifiable information that users had likely supplied themselves, including dates of birth, genders, contact information and home addresses.
An ocean of data, including chat logs, generated images and videos were left exposed. More: https://t.co/6lo9NlpXtF pic.twitter.com/TaWeoFAp4T
— Cybernews (@Cybernews) October 21, 2026
Researchers also found more than 26,000 ZIP files containing conversation summaries, user inputs, AI-generated outputs and files attached to conversations. Other exposed documents covered subjects including large language model development research and material that appeared to have been uploaded for private analysis.
The archive therefore contained more than individual chatbot exchanges: it also included files and documents that users or businesses may have expected to remain private.
Sex Doll Maker's Support Chats Exposed
Among the most distinctive findings were multiple logs from a customer-support chatbot belonging to a prominent sex doll manufacturer. Cybernews did not name the company, and its report did not establish that every customer interaction handled by the manufacturer was exposed.
'What we found was that some companies used the service for enterprise tasks. For example, we found multiple chat logs from a customer support chatbot owned by a prominent sex doll manufacturer,' the Cybernews research team said.
The discovery highlights the potential privacy risks when businesses use third-party AI services to handle customer interactions and other work. However, Cybernews did not confirm that the exposed information had been stolen or misused.
The Risk of Third-Party AI
302.ai acts as an intermediary between users and AI model providers. According to Cybernews, its customers can access major large language models, including ChatGPT, Claude, Grok, DeepSeek and Gemini.
The exposure occurred in 302.ai's storage infrastructure and does not establish that the AI model providers themselves suffered a breach.
The distinction matters because users may interact with AI through third-party services rather than directly through the companies that develop the models.
'Both individuals and businesses increasingly trust AI agents with sensitive information, be it personal details or business operation info,' the Cybernews research team said.
It added: 'Therefore, not only the primary AI companies, but also businesses that handle these agents bear the responsibility to handle whatever information that can occur in these chats as securely as possible.'
The incident raises questions about how third-party AI platforms store conversation histories, protect uploaded files and restrict access to customer data. The investigation does not, however, establish which specific security failure caused the exposure or what security measures 302.ai had in place.
How Scammers Could Exploit the Data
Cybernews found no evidence that malicious actors had accessed or downloaded the exposed files. However, if obtained by criminals, chat logs and attachments could reveal personal details that might make phishing or other social-engineering attempts more convincing.
The investigation did not establish that the information had been used in such attacks. These are potential consequences of the exposure, not confirmed cases of criminal exploitation.
Timeline of the Exposure
Cybernews provided the following timeline for the incident:
- 2 August 2026: Researchers discovered the exposed data.
- 11 August 2026: Initial disclosure.
- 8 September 2026: Researchers observed that the exposed storage had been closed.
Cybernews said the data was taken offline after its researchers contacted 302.ai. Its published timeline places the discovery on 2 August and the initial disclosure on 11 August, while the storage was not observed to be closed until 8 September.The publication said it had sought comment from the company and would update its report if a response was received.
302.ai is developed by Univerads Technology Limited, a Hong Kong-based company founded in 2020, according to Cybernews.
The report leaves important questions unresolved, including whether affected users were notified, whether further security changes were made and whether any unauthorised party accessed the files. A response from 302.ai would help clarify these points.
