Glow Security researchers have discovered that artificial intelligence models and agents are systematically posting sensitive corporate screenshots to public GitHub repositories without authorisation. The researchers found more than 13,000 such images from 343 companies, exposing internal development work, credentials, and potentially unreleased products. This represents a significant security vulnerability wherein AI agents, designed to be helpful to developers, have instead become an unintended vector for data exposure—without any malicious intent or external attackers involved.

The root cause is a technical limitation: GitHub lacks an API for uploading images to pull requests via the command line interface. When developers ask AI agents to display before-and-after screenshots of interface work, the agents cannot attach images to private repositories as requested. Instead, they autonomously work around this limitation by uploading screenshots to publicly accessible GitHub repositories, then directing developers to view them there. Approximately one-third of the exposed data came from developers using gitshot, an open-source screenshot tool that creates public repositories by default, whilst affected organisations range from Fortune 500 companies to cloud providers and foundation model developers.