15:47Machine-translated from Azerbaijani by AIRead in Azerbaijani
AI agents exposed more than 13 000 confidential screenshots online
AI agents operating without oversight have caused data leaks at more than 300 organizations. In total, more than 13 000 confidential screenshots were exposed on the open internet. The victims include several companies on the “Fortune 500” list and leading AI laboratories. The incident was uncovered by cybersecurity company “Glow” in its “PixelLeak” report.
During software development, screenshots are added to code update requests to show changes made to the interface. Known among developers as PRs, these requests are used to review changes. Although images can be added through “GitHub”'s graphical interface, the command line used by bots does not support this in a private code repository. As a result, AI agents uploaded the images to a public code repository and then added links to those files to the private request. At approximately 33 percent of the affected companies, developers used a command-line tool called “gitshot” to add images. The files were easy to find because the tool added the “_gitshot” tag to the image names. In 93 percent of cases, the images were found not in the company's “GitHub” account but in a public code repository managed under a developer's personal username. As a result, files belonging to private projects remained accessible to outsiders.
In one case, lengthy instructions given to an AI agent made the leak even larger. The agent saved the method of uploading images to a public repository as a permanent rule and used it in subsequent programming tasks as well. As a result, images of features that were not supposed to be shown to the public until several months later were exposed. The agent's notes revealed that it had created a separate public repository for PNG images.
The leaked files included images of internal and unreleased software, company and customer data, and financial documents. Screen recordings of a money transfer interface were also found among the public files. The leak put not only software code but also real-world operations and customer privacy at risk. Because the images appeared normally in the private request, the companies may have remained unaware of the problem. To reduce the risk, “Glow” recommends prohibiting employees from creating code repositories in their personal accounts. The accounts of former employees and the code they managed should also be reviewed. AI tools used without the IT department's knowledge should be restricted, and software libraries and instructions given to agents should be carefully reviewed. These steps could prevent AI agents from uploading confidential files to the open internet.
During software development, screenshots are added to code update requests to show changes made to the interface. Known among developers as PRs, these requests are used to review changes. Although images can be added through “GitHub”'s graphical interface, the command line used by bots does not support this in a private code repository. As a result, AI agents uploaded the images to a public code repository and then added links to those files to the private request. At approximately 33 percent of the affected companies, developers used a command-line tool called “gitshot” to add images. The files were easy to find because the tool added the “_gitshot” tag to the image names. In 93 percent of cases, the images were found not in the company's “GitHub” account but in a public code repository managed under a developer's personal username. As a result, files belonging to private projects remained accessible to outsiders. In one case, lengthy instructions given to an AI agent made the leak even larger. The agent saved the method of uploading images to a public repository as a permanent rule and used it in subsequent programming tasks as well. As a result, images of features that were not supposed to be shown to the public until several months later were exposed. The agent's notes revealed that it had created a separate public repository for PNG images. The leaked files included images of internal and unreleased software, company and customer data, and financial documents. Screen recordings of a money transfer interface were also found among the public files. The leak put not only software code but also real-world operations and customer privacy at risk. Because the images appeared normally in the private request, the companies may have remained unaware of the problem. To reduce the risk, “Glow” recommends prohibiting employees from creating code repositories in their personal accounts. The accounts of former employees and the code they managed should also be reviewed. AI tools used without the IT department's knowledge should be restricted, and software libraries and instructions given to agents should be carefully reviewed. These steps could prevent AI agents from uploading confidential files to the open internet.