Coding agents left more than 13,000 private screenshots where anyone could fetch them. The files came from over 300 organizations, including several Fortune 500 companies and a frontier AI lab.
Tom's Hardware is summarizing PixelLeak, a report from the security firm Glow. Both "more than 13,000" and "over 300" are floors. Their quotient is not an average number of images per company. Either count can be higher, so the true average can sit above or below 13,000 divided by 300. The pictures reportedly show unreleased software, client data, financial information, and recordings of a money-movement screen.
A private pull request cannot show the image, so the agent made a public one
People attach screenshots to pull requests so a reviewer can see what changed. GitHub's website lets a human do that. An agent on the command line, Glow says, cannot attach an image inside a private repository the same way. The agents' fix was to keep the pull request private and link the picture to a file in a public repository.
Glow quotes the reasoning one agent wrote down. A private repo "cannot render images" in the description because GitHub's image proxy "fetches anonymously," so a file kept private "shows up broken for reviewers." The agent made a public repo, pinned two screenshots to a commit, and called the problem solved. That is a transcript, not a guess about motives.
About a third of the affected companies were using gitshot, a command-line screenshot tool, for the same limitation. Glow says those files are easy to find: search for the tag "_gitshot." "About a third" is not a counted 33 percent, and it should not be turned into a headcount by multiplying 300.
In 93 percent of cases the images lived in repositories under the developer's username, not the company's GitHub organization. Tom's Hardware says "cases." That is a share of cases in Glow's set, not automatically 93 percent of the 13,000 images. One agent skill, a long prompt that teaches the bot a habit, spread the workaround until agents used it on every ticket, including features still months from release.
The tools that catch this class of leak are still sold to a short list of buyers. Newell's rule about hackers assumes someone is trying to break in. Here the agent was trying to help a reviewer.
Glow's public write-up, as Tom's Hardware relays it, does not name the Fortune 500 firms or the lab. The gitshot tag is searchable. How many of those 13,000 files are still online is the number the report does not give.
A skill file turned a workaround into a habit
The leak is not a stolen database. It is a habit. Humans attach images in the GitHub interface. Agents, lacking that command for private repositories, hosted the PNG elsewhere and left a link. Glow says that in one case the workaround was written into an agent skill, and then used on every development ticket. Features "months away from public release" were in those frames. Months is not a count of features. It is the lead time the report claims was exposed.
The money-movement recordings are the sharpest category in the list, beside financial data and client information, and they are not given a separate tally. A reader cannot say what fraction of the 13,000 is a payroll screen. The 93 percent figure answers a different question: whose repository, not what was on the screen. Searchable tag, unnamed companies, floors rather than totals. That is the whole public shape of PixelLeak as Tom's Hardware prints it. The command-line gap is the cause Glow calls simple. GitHub shows a human a button. It does not show an agent that button for a private repository. The agents routed around the missing button. The companies got the screenshots in the pull request and, unless someone asked, no notice that the file was public.