Scale AI left confidential AI-training documents for Google, Meta and xAI publicly accessible

Business Insider reported that Scale AI, the data-labelling company central to training pipelines at several frontier labs, had left confidential client documents publicly accessible through unsecured, shareable Google Docs links. The outlet said it reviewed thousands of pages across at least 85 documents, some of which remained editable by anyone with the link, tied to Scale AI’s work for Google, Meta and xAI.

The exposed material included, for Google, internal instruction manuals describing specific weaknesses in its Bard chatbot — such as difficulty handling complex questions — along with guidance for the contractors Scale AI employed to help fix them. Meta’s documents included labelled audio clips and standards for chatbot speech training. xAI’s exposure covered details of at least ten generative-AI projects, including one, “Project Xylophone,” built around roughly 700 prompts intended to improve conversational quality. Separate spreadsheets exposed Scale AI contractors’ private email addresses, pay disputes, and internal performance ratings, including workers flagged for suspected cheating.

Scale AI said it had disabled the ability to publicly share documents from its managed systems and opened an investigation after being notified. The incident illustrated a security risk specific to the AI supply chain rather than to model training itself: large volumes of client instructions, evaluation criteria and raw labelled data pass through data-labelling vendors’ ordinary office software, and a permissions mistake there can expose competitively sensitive information — including, in this case, one lab’s documented assessment of a rival’s product weaknesses — without any breach of the AI systems themselves. It came roughly two weeks after Meta announced a multibillion-dollar investment in Scale AI, intensifying scrutiny of the company’s data-handling practices.