AI aggregator 302.ai has exposed 22.72 terabytes of user data, including private chatbot conversations, personal information, multimedia files, and corporate documents, following a security lapse involving publicly accessible cloud storage.
The exposure was uncovered by cybersecurity researchers at Cybernews, who found approximately 17.6 million files in an Azure storage bucket associated with the platform. The discovery included more than 144,000 chat logs, raising concerns about the privacy of individuals and businesses using third-party AI services.
302.ai provides access to multiple artificial intelligence models through a single platform, including ChatGPT, Claude, Grok, DeepSeek, and Gemini. Such services allow users to switch between different models without managing separate interfaces, but they also introduce another company that may handle prompts, uploaded files, and conversation histories.
According to the researchers, the exposed storage contained nearly 9 terabytes of images, 7.7TB of video files, and 5.6TB of audio recordings. The archive also included documents and compressed files containing AI conversation data.
The most concerning discovery involved more than 144,000 chat logs stored in JSON format. These records included user prompts, AI-generated responses, model selections, token costs, timestamps, and other metadata. Some conversations reportedly contained personally identifiable information, including birth dates, gender details, contact information, and home addresses.
The incident also affected enterprise-related information. Researchers found chat records associated with business activities, including customer support conversations from a chatbot operated by a prominent manufacturer of sex dolls. Other exposed documents reportedly included AI development research and private files that users had submitted for analysis.
More than 26,000 ZIP archives were also found in the exposed storage. These files contained conversation summaries, user inputs, AI-generated outputs, and attachments associated with chatbot interactions.
The discovery was not reported as a confirmed cyberattack. Instead, the researchers suspected that the information had been exposed accidentally because of a cloud storage configuration problem.
The timeline provides additional context. According to reporting by TechRadar, Cybernews discovered the exposed data in early August 2026 and subsequently contacted the company. The storage was reportedly secured by September 8, although the exact date when the exposure first began remains unknown.
Cybernews said it had no evidence that malicious actors accessed or downloaded the files. However, the researchers warned that the absence of confirmed misuse does not establish that nobody else discovered the publicly accessible information before it was secured.
The company behind 302.ai, Univerads Technology Limited, is based in Hong Kong and was founded in 2020. As of the latest reporting, the company had not publicly responded to the researchers’ request for comment.
The exposed information could have significant consequences if obtained by cybercriminals. Private AI conversations may reveal personal circumstances, workplace activities, business strategies, or confidential documents. Attackers could use such details to create highly personalized phishing messages, impersonate trusted contacts, or target organizations with convincing social engineering attacks.
The incident also highlights a security risk that users may overlook when choosing AI tools. A conversation submitted through an aggregator may be handled by the intermediary as well as the underlying AI model provider. Users should therefore review the intermediary’s privacy policy, data retention practices, and security protections before sharing sensitive information.
Businesses should assess third-party AI services before allowing employees to upload customer records, internal documents, source code, or confidential company information. They should also establish clear rules for using AI tools and investigate whether any credentials or sensitive data may have been included in exposed conversations.
If this article helped you, please consider supporting our work. Every small contribution keeps Abijita.com independent and running.
Although the reported storage exposure has been closed, questions remain about how long the data was publicly accessible and whether any unauthorized parties obtained copies. The incident serves as a reminder that protecting AI conversations requires more than securing the models themselves. Every platform that stores or processes user information must also protect that data against unauthorized access.
