// TOM'S HARDWARE US — HARDWARE & GADGET
AI agents inadvertently leak 13,000+ internal screenshots from organizations
When you purchase through links on our site, we may earn an affiliate commission. Here’s how it works.
There's a private information leak most days, usually by way of misconfigured services or nasty security bugs. Sometimes, though, users will readily hand over private information without being aware of it. That's the case for over 300 organizations, including several Fortune 500 companies and a frontier AI lab, who collectively had 13,000+ private screenshots exposed — all thanks to their development AI agents being arguably too good at their jobs and performing them with little human oversight.
Besides showing pictures of internal and pre-release software, the leaked screenshots reportedly include corporate and client information, financial data, and even screen recordings for a money-movement interface. The report, called PixelLeak, comes from endpoint security firm Glow, and it details how the leaks happened. The root cause is surprisingly simple and likely to induce a forehead slap.
It's become customary in development work related to UI and UX (and other categories) to include screenshots showing previews or before/after comparisons of tweaked features, for review purposes. Said images travel as attachments to the respective code changes, also known as "pull requests" (PRs) in dev parlance.
When using GitHub (and potentially other code repository services), humans see a graphical interface for easily attaching an image to a PR. Meanwhile, bots are limited to using the command-line interface, which currently does not have that feature available for private repositories.
As the efficient and smart agents they are, the clankers came up with a simple solution: publish the PR as usual to the private repository, and include an image placeholder linking to a file that's hosted in a public repository instead. There, problem fixed! The user is happy and likely has no idea what happened unless they notice the problem somehow and start asking the bot some hard questions.
Glow says that in about a third of the affected companies, developers were using gitshot, a command-line tool for attaching screenshots, used in this instance to overcome the aforementioned limitation. Searching for images attached by the tool is quite easy, as one needs but look for the "_gitshot" tag. The report also indicates that in 93% of cases, images were found in repositories under direct control of the developer's username, rather than being tied to the company's GitHub account.
In one particular case, an agent skill (essentially long-winded prompts instructing a bot how to do something) was also an indirect source of leaked information. The agents started incorporating the image hosting workaround as a skill, and after a while, many of them were using it for every development ticket, leading to leaks of information about features months away from public release.
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Glow provided sample reasoning output from an agent, explaining precisely why it uploaded the screenshots publicly: