Itdaily - AI coding agents leak more than 13,000 internal screenshots via GitHub

AI coding agents leak more than 13,000 internal screenshots via GitHub

Illustration of the GitHub logo, the Octocat, in green on a dark grid background.

Glow Labs reveals how AI coding agents inadvertently posted more than 13,000 internal screenshots publicly on GitHub. Large enterprises in cloud, fintech, healthcare, and government also appear to be affected.

Developers are increasingly relying on AI agents to wrap up their workday: summarizing changes, adding screenshots, and submitting everything for review. This convenience has an unexpected downside. Research from Glow Labs shows that these agents are proving visual changes en masse by publishing screenshots outside the secure development environment.

In the PixelLeak study, Glow Labs identified more than 13,000 internal images that were openly available on GitHub, spread across over 300 organizations. These include pre-release screenshots of unreleased features and sensitive customer data. Security teams noticed nothing, as the leaks often occurred via personal accounts and public side-repositories.

How AI agents make screenshots public

Every case investigated started with a seemingly innocent request: show with screenshots that a UI fix works. The agents hit a limitation of GitHub: the integrated image hosting is tailored for humans in the browser, not for text-based CLI agents. Linking images directly to a private pull request failed, leaving the review incomplete.

To bypass this, the agents “devised” a technical workaround: they hosted the images in an adjacent public repository, which was easily accessible to reviewers. The functional objective was achieved, but the security implications were completely ignored. As a result, internal screens from private repositories suddenly appeared in world-readable GitHub projects, out of sight of traditional organization-wide security controls.

Scale

The leak affects more than 900 code repositories at enterprises with up to more than 100,000 employees, active in sectors such as cloud, healthcare, fintech, government, frontier AI, and even AI security. Several involved parties are Fortune 500 companies. Anyone with the correct URL could download and analyze the images, including competitors or malicious actors.

In one case, a developer at a large manufacturer made an AI agent responsible for testing an internal billing screen. The agent autonomously created a public repository in the employee’s personal GitHub account and posted screenshots there containing billing data from a utility company. Because this happened outside the company’s GitHub organization, the incident remained invisible to internal security tools until Glow Labs explicitly informed them.

Role of gitshot

About a third of the affected organizations used gitshot, a small open-source tool that publishes screenshots for code reviews. AI agents discovered this tool independently in development environments and utilized it to bypass the GitHub CLI limitation. Images ended up under the _gitshot tag, easy to find for those who know where to look.

Glow Labs found more than 100 public accounts leaking internal development work in this way. This included an AI frontier model company, a financial service provider with images of treasury and settlement consoles, and a payments company where multiple employees had their own gitshot repositories. At one enterprise software player, one agent solution even became standard practice: more than a dozen agents uploaded over a thousand screenshots and screen recordings of unreleased features within a single week.

Promises of security

AI agents are increasingly succeeding in gaining unauthorized access to non-public data. They go to great lengths to ensure their tasks succeed, even if it comes at the expense of security and privacy. More and more incidents are coming to light where AI agents step outside the boundaries of their test environments. Consider the breach of OpenAI’s AI agents into Hugging Face, or the recent incident where AI agents penetrated the Canadian government’s medical database.

Promises of security are becoming increasingly difficult to uphold. This is not only due to AI agents but also features like Windows Recall. This tool takes screenshots of what happens on your PC at regular intervals. However, cracks in that security system have also recently been shown.