Scale AI reported project materials for clients such as Meta and Xai, followed by a Business Insider report, which included thousands of sensitive files stored as Google documents published in the link.
The company, which uses human gig workers to improve Big Tech's latest AI model and receives $14 billion investment from Meta, is scrambling this week to secure files.
This prevented the worker team from opening training documents temporarily. Thousands of scale AI files previously reviewed by BI, previously published, are now private.
“What's going on is a knee-jerk response to being in the headline,” Stephanie Kurtz, regional director of cybersecurity firm Trace3, told BI. She said it should have been done in the first place, locking down the documents and invited the right users.
By Wednesday, the team had resolved many of the document access issues, one worker said. Another contributor, said to be another, is currently granted separate access to the document.
BI first flagged Google Docs and expanded its AI in its June 13th article on how it showed Google using ChATGPT to improve AI chatbots. BI also encountered public-scale AI documents in previous reports on how Xai and Meta train their latest AI models.
At least 85 individual Google docs, including thousands of pages, remained fully accessible until BI published an article Tuesday focusing on security issues created by the practice.
BI reported that Scale AI has kept thousands of project documents open related to working with clients, including Google and Meta, making them accessible to anyone with the link. Some documents also included contact information for numerous scale AI workers, some of whom were surprised they found that details were accessible when BI contacted them.
Scale AI is an efficient way to share information with over 240,000 contractors, and we use public Google Docs on a daily basis to track work for our well-known customers. BI found that these documents often contain sensitive information about how workers train AI models for large high-tech clients. Multiple AI training documents reviewed by BI are labeled “confidential” and accessible to anyone with the link.
After AI reductions, one contractor explained the “site-wide” issue on Tuesday when accessing project materials. Another said many teams' jobs are based on halting due to new restrictions and may even lose access midway through important presentations.
“We're basically cold here,” the contractor said.
Scale AI told BI on Monday that it was conducting a thorough investigation, nullifying the user's ability to publish documents from the scale's system.
Repeated the statement in this article and did not comment further on the specific changes they made to expand document security for AI.
“We take data security seriously,” said an AI spokesperson. “We continue to be committed to robust technical and policy protection to protect sensitive information and are constantly working to strengthen our practices.”
There is no indication that Scale AI has suffered from a data breach. Cybersecurity experts told BI that the practice could make the company vulnerable to hacking.
The lockout of the document was a bit more whiplash for scale AI contractors influenced by Mega Investment and the decision to hire CEO Alexandre Wang.
After the transaction announcement, Google stopped several projects with Scale AI. Openai and Elon Musk's Xai have previously reported that they are suspending projects at scale, with one small investor saying they are selling the rest of the stock in the startup.
Many contractors have found that some of their projects have been suspended. Scale AI sent memos to contractors announcing meta investments, but many workers said they were left in the dark about clients who suspend the project primarily without prior warning.
Meta declined to comment.
Any hints? Please contact Charles Rollet by email crowdet@businessinsider.com Or with signal and whatsapp 628-282-2811. Please contact Hullangley by email hlangley@businessinsider.com Or Hughlangley.01 signal. Use your personal email address and unprocessed devices. Here's a guide to sharing information safely.
