PixelLeak Exposes Sensitive Data in More Than 13,000 AI-Agent Screenshots

Cybersecurity researchers at Glow Security found more than 13,000 screenshots in public GitHub repositories created by AI agents. The images came from more than 300 companies and, in some cases, showed information about sensitive corporate software projects. The researchers call the findings “PixelLeak.”

According to Glow Security, developers asked agents to provide before-and-after screenshots so reviewers could check visual changes. Because coding agents use text-based command-line tools, they could not attach images through GitHub’s pull request interface. Some agents worked around the limitation by uploading screenshots to adjacent public repositories, without accounting for the exposure risk.

Glow said it identified 343 organizations leaking sensitive information this way, with more than 900 code repositories affected. One case involved a manufacturer with more than 100,000 employees. An agent uploaded screenshots of work on an internal billing screen to a public repository in a developer’s personal GitHub account. Glow said the images included billing records for a utility company involved in the interface fix.

The researchers also linked some cases to gitshot, an open-source tool that publishes screenshots for code reviews. They said about a third of affected organizations had developers using it, and that more than 100 public accounts exposed internal development work through images tagged _gitshot. Glow said it contacted the organizations it identified and warned that others could also be affected.