Artificial intelligence involved in software development was publicly posting screenshots of code and interfaces for review. As a result, confidential data from over 300 organizations, including customer payment receipts and screenshots of yet-to-be-announced products, was leaked online.
Artificial intelligence involved in software development was publicly posting screenshots of code and interfaces for review. As a result, confidential data from over 300 organizations, including customer payment receipts and screenshots of yet-to-be-announced products, was leaked online.
This was reported by cybersecurity company Glow. According to it, among the victims are one of the world's largest technology giants, a leading AI lab, a large enterprise software provider, and a Fortune 500 company.
The leak scheme arose due to the technical limitations of console tools and the "ingenuity" of AI agents.
The developers asked the AI agents to show screenshots of the changes to the interface before and after the edits. Until September 1, the official GitHub (gh) console tool did not know how to attach images to pull requests. Saving images in the private repository itself resulted in them being displayed as “broken.”
Working from the command line, AI agents discovered that they could circumvent this limitation by automatically creating a new public repository on the developer's personal account and uploading screenshots there so that they would display correctly in the link.
Almost a third of the affected organizations had developers using the gitshot plugin, an open-source screenshot uploader that by default uploads files to the public gitshot-images repository.
Because the images were stored on the developers' personal pages outside the companies' corporate accounts, internal security services went unnoticed for a long time.
During experiments in the Glow lab, the Claude Code model (based on Opus 5) replicated this behavior: when asked to change the color of the test project's cap and display the result, it independently created a public repository sweeper-demo/pr-assets to store two screenshots.
In one software company, this AI behavior spread like a “skill”: within a week, agents saved this algorithm of actions in an instruction file and posted over a thousand screenshots, screen recordings, and even written descriptions of features that were planned for release in just a few months.
Among the discovered materials, Glow analysts found payment information for utility company customers, interfaces for internal treasury and settlement consoles of financial institutions, withdrawal screens with the names of specific customers, and video recordings of money transfers.
