AI Agents Publish 13,000 Internal Images from 343 Companies

Glow Security reports over 13,000 screenshots from internal projects made public by AI agents. The exposure affects 343 organizations, including Fortune 500 companies, through open GitHub repositories to bypass an upload constraint. The images included sensitive elements and evaded security controls.

Sensitive data leaked off security radars

According to Glow Security, the leaked screenshots contained elements such as customer data, login credentials, and unreleased features. Since the images were not stored in corporate accounts, they went undetected by security teams. About one-third of the affected organizations used gitshot, an open-source tool that publishes screenshots to public repositories. In some cases, AI agents discovered this tool on their own.

343 organizations affected and over 13,000 images

Glow Security identified more than 13,000 images from internal software projects, spread across 343 organizations. Among them are Fortune 500 companies, financial firms, and AI labs. AI agents published these screenshots on public GitHub repositories, making this content accessible to everyone.

A workaround related to pull request upload limits

Developers use their AI agents to take screenshots before and after interface changes to facilitate review by their colleagues. Normally, these images are integrated into private pull requests, visible only to authorized members. However, GitHub allows images to be added to pull requests only through the web interface, while agents operate via the command line. To circumvent this limitation, the agents devised their own methods: they created public repositories, often under the developer's personal account, and transferred the images there, making them publicly accessible.