In the world of artificial intelligence, where data confidentiality is critical, Scale AI is now facing serious questions about its internal security practices. It has been uncovered that the fast-growing startup, which works with clients like Meta, Google, and Elon Musk’s xAI, has been leaving sensitive internal documents accessible to anyone with a link—using public Google Docs.
Thousands of confidential documents related to AI training projects were available online without restriction. Some files carried labels like “confidential” and contained information about major clients’ AI initiatives, internal project guidelines, and even personal data about Scale AI’s own workforce.
Sensitive Client Projects Exposed
Among the most troubling discoveries were documents showing details of Google’s work to improve Bard, its struggling chatbot product. The leaked docs included step-by-step instructions for Scale contractors on how to help fix Bard’s issues with complex questions. Other confidential Google files, labeled as such, were also found publicly accessible.
Similarly, internal documents tied to xAI’s generative AI projects, including one called “Project Xylophone,” were left unsecured. These included a list of training prompts ranging from casual conversations about zombie apocalypses to more technical topics like plumbing. Meta’s files, also marked confidential, exposed AI training materials, example audio files, and internal guidelines designed to make Meta’s chatbots more conversational and emotionally engaging.
Contractors working on these projects told that it was often obvious which client they were working for—even when client names were masked. In some cases, Google’s own logo appeared in presentations. In others, the AI models would reveal their true identity just by being asked.
Contractor Privacy at Risk
The security gaps weren’t limited to client data. It was found that spreadsheets listing the names and personal Gmail addresses of thousands of Scale AI contractors were publicly accessible. Some documents contained sensitive performance evaluations, with categories such as “high quality” and “suspected of cheating.” One spreadsheet even titled “move all cheating taskers” flagged hundreds of individuals for “suspicious behavior.”
In other cases, contractors who had been mistakenly banned were listed in public docs. Some files went as far as including payment records, disputes over wages, and other HR-related notes.
A Widespread, Longstanding Practice
Scale AI’s reliance on public Google Docs appears to be more than a one-off mistake. Multiple contractors said this was a routine practice, driven by the sheer size of Scale’s contractor base — estimated at around 240,000 workers.
While using public links made it easier for Scale to quickly distribute instructions across its freelance workforce, it came at the cost of exposing internal data to potential bad actors.
The company’s internal platform does include identity verification steps for contractors, but links to important training documents remained freely accessible to anyone with the right URL — no verification required.
“The whole Google Docs system always seemed incredibly janky,” one former contractor said.
Some freelancers even reported retaining access to old projects long after their work had ended, including documents still being updated with new client instructions.
Expert Warnings on Cybersecurity Risks
Cybersecurity experts say the risks from such practices are clear and alarming. Joseph Steinberg, a cybersecurity lecturer at Columbia University, described the situation bluntly: “Of course it’s dangerous.” Public exposure of contractor lists could make it easy for hackers to launch social engineering attacks, impersonating staff to gain deeper access to Scale AI’s systems.
Another risk involves hackers inserting malicious links into editable documents. Stephanie Kurtz, a director at cybersecurity firm Trace3, stressed that leaving files publicly editable is especially risky. “Putting it out there and hoping somebody doesn’t share a link is not a great strategy,” she said.
Scale AI’s Response
In response to these findings, Scale AI said it had already launched a full investigation and disabled the ability for users to publicly share documents from Scale-managed systems.
“We remain committed to robust technical and policy safeguards to protect confidential information and are always working to strengthen our practices,” a Scale AI spokesperson said.
Meta declined to comment, while Google and xAI did not respond to inquiries.
This incident comes at a sensitive time for Scale AI. The company is fresh off a $14.3 billion investment from Meta and is trying to reassure Big Tech clients that it remains a neutral, secure, and independent partner. Last week, Scale posted a blog promising its clients that “ensuring customer trust” is a top priority.
Yet with these findings raising serious concerns, the real test for Scale AI now lies in how quickly and transparently it can fix these lapses before further damage is done.
Prepared by Navruzakhon Burieva
Leave a Reply