Elon Musk’s Grok has exposed more than 35,000 applications to anyone on the interwebs.
Security boffins at Claroc discovered that 35,578 applications created using Grok’s Build tool could be opened without signing in. The exposed material included internal company documents, medical information, employee passwords and sales records. Among the victims was SpaceXSI, the company behind Grok, whose staff had left 24 applications accessible to strangers.
Grok Build allows users to create working applications by describing what they want in ordinary language. The system writes the code, assembles the application and provides a way to publish it online. Unfortunately, it seems that teaching software to build applications is considerably easier than teaching it to keep its customers’ secrets.
According to Claroc, researchers identified 43,601 application addresses under Grok’s grok.me domain and tested 43,551 of them. Of the 37,571 applications still operating, 94.7 per cent opened without authentication. Only 1,993 required visitors to sign in, suggesting that making applications publicly accessible had become a popular choice among people who might not have understood what they were publishing.
The security problem stems from how Grok handles published applications. Each receives a web address made from several apparently random words. While these combinations might be difficult to guess, the addresses appear in commercial passive DNS databases whenever somebody looks them up. Anyone with access to those databases can discover the supposedly obscure links without breaking into anything.
Claroc found that the applications contained considerably more than experimental games and calculators. Among the exposed material was a hospital staff directory containing information about more than 3,500 employees, an intensive care patient’s medical briefing and a contractor’s timesheet application containing 18 employee accounts with passwords stored in plain text. Presumably, the person responsible for the last one thought password security was something that happened to other people.
Other applications exposed information from construction companies, utilities, defence contractors and transport businesses. Claroc identified 30 organisations whose internal material was accessible, including six Fortune 500 companies. The researchers estimated that between 650 and 800 applications contained private information serious enough to warrant considerable concern.
The most embarrassing discovery involved SpaceXSI itself. Researchers found an internal onboarding document for a new enterprise sales employee, complete with details of company systems, internal communications channels and a 30-day training plan. The sales territory map listed 594 strategic accounts, and the advertising proposals and campaign reports covered customers in several countries.
One advertising proposal was intended for a US federal agency and marked confidential, although its security arrangements suggested that the word was largely decorative. Another collection of internal material included live advertising account boards and sales targets. SpaceXSI subsequently removed the 24 applications belonging to its employees after the researchers raised the alarm.
Claroc wanted to establish whether the trouble extended beyond people accidentally publishing private information. It created two test applications with fictitious data and examined the security controls Grok generated. One application allowed anonymous visitors to add records, while another supposedly protected customer information with Google authentication. Unfortunately, the second application accepted an uninvited Google account and allowed its owner to view, modify and delete customer records.
The researchers stressed that they conducted these experiments on applications they created themselves. They did not attempt to modify the information belonging to other organisations. However, their tests demonstrated that an application which appears protected by a login screen might still allow outsiders to access information if the software fails to check whether they are authorised.
SpaceXSI defended its publishing arrangements, arguing that applications remain private until their creators deliberately publish them. The company said the process involves several steps and that employee accounts use private access settings by default. Claroc accepted that publishing requires user action but found that its own paid account preselected public access. Apparently, clicking through several screens is supposed to make an insecure decision more respectable.
The researchers disclosed their findings to SpaceXSI on 8 November and supplied further details the following day. By 10 November, the company had removed its employees’ exposed applications, but Claroc said those belonging to other organisations remained available. The security outfit began contacting affected companies itself after receiving no indication that SpaceXSI intended to notify the owners.
The affair is another warning about businesses embracing AI-generated software without considering what happens after the code has been written. Claroc found that employees could create and publish internal applications without involving their IT departments, leaving potentially sensitive information outside established security controls. The researchers recommended private access settings by default, stronger authentication, and warnings when applications containing databases are published without protection.
