AI coding agents posted 13,000 internal screenshots to public GitHub repos

Glow Security says AI coding agents have uploaded more than 13,000 internal screenshots from 343 organisations to public GitHub repositories, The Decoder reported. The set includes Fortune 500 companies, financial institutions and an AI research lab. Tom's Hardware, which also covered the finding, put the number of affected companies at more than 300.

The agents were solving a small problem. GitHub only lets users attach images in a browser, not from a command line, so the agents created public repositories, usually in a developer's personal account, and dropped the images there. Glow Security says the images showed customer data, login credentials and details of unreleased product features.

About a third of the cases involve GitShot, an open-source command-line tool by developer Vipul Gupta, according to Glow Security. By default it creates a public repository called gitshot-images. Its documentation warns users not to upload sensitive information, a warning Gigazine says the agents ignored. Glow Security stresses the leak is not limited to one tool or model.

The leak was hard to see. Because the images sat in personal accounts instead of company ones, security teams did not notice, The Decoder reported. In one case an agent at a manufacturer posted billing screenshots containing authentication data to a personal account. Glow Security's Omer Singer said agents lack the common sense to avoid doing things they should not.

None of the reports we read describes a company response, and Glow Security's own figures have not been independently checked. Glow Security's advice, as relayed by Gigazine, is that organisations should restrict not only what agents are for but also which tools they can reach and what they may disclose.