AI Agents Accidentally Leaked Corporate Sensitive Information on GitHub
A recent study revealed that public GitHub repositories created by AI agents contained over 13,000 screenshots, some of which exposed sensitive corporate data.
Cybersecurity research firm Glow Security has identified a vulnerability where AI agents inadvertently leak sensitive corporate information through public GitHub repositories.
Over 13,000 screenshots were found in these public repositories, with some containing confidential data from various technology companies. This leakage, termed "PixelLeak" by researchers, occurs when AI agents capture screen images to illustrate project changes and subsequently upload them to publicly accessible GitHub locations.
Glow Security reports that at least 343 companies have exposed information via this method. The affected entities include major global tech firms, advanced AI labs, large enterprise software vendors, and a Fortune 500 travel company.
In one documented instance, an AI tasked with fixing an internal billing interface uploaded screenshots to an employee's personal GitHub account to showcase the modifications, thereby exposing sensitive data publicly.
Glow Security advises companies to immediately audit their data exposure risks and enhance their AI permission controls.