Public GitHub Repositories Leak 13,000 Images via AI Coding Screenshots

Exposed Screenshots Become a Code-Review Side Channel

A security disclosure on October 1, 2026, said more than 13,000 internal images from 300+ organizations were found in public GitHub repositories after AI-assisted code reviews captured screenshots and uploaded them alongside code changes. The researchers, named Glow in the disclosure, described the leak as a side effect of autonomous developer agents that take and store screen captures as part of their review workflow, turning ordinary review artifacts into sensitive data exposure. Two outlets reported the same finding: framed it as "AI Coding Agents Exposed 13,000 Internal Images," while repeated the headline through the lens of "Shadow AI" inside enterprises.

Both stressed that the screenshots were public, indexed on GitHub, and originated from real organizations using AI coding tools in production.

What This Breaks for Engineering Teams

The practical fallout is that AI coding agents now touch data the developer never intended to commit. A code review that includes a screenshot of an internal dashboard, a configuration screen, or a customer record will, with some agent configurations, push that image into version control where it becomes searchable to anyone who knows the pattern. Security teams should treat AI-driven code reviews the same way they treat pasted logs: scrub for screenshots before merge, gate public repositories, and audit existing repositories for stray image attachments introduced by agents.

Two practical defenses any developer can run today are visible in the linked guides: the MIME Type Lookup helps confirm whether a stray file in a repository is in fact an image, and the Rotate Multiple Images Locally Without Uploading walk-through demonstrates that image handling can stay on-device, which is the same principle teams should apply to agent screenshot flows.

Copilot's Plugin Surface Opens to External Authors

On the same day, Microsoft's developer blog introduced Work IQ Developer Tools (WIQD), a path for turning a business problem or inefficient workflow into a plugin that extends what Microsoft Copilot can do. The framing matters for working developers: Copilot is shifting from a chat surface to a host platform, and external plugin authors will become responsible for the actions Copilot takes inside enterprise data. Teams that already maintain internal automation now have an on-ramp to ship that logic as a Copilot extension rather than a standalone script.

AI Coding Tools Get an API-First Deployment Path

js hosting product that lets developers and AI coding tools deploy and manage web apps through an API. The announcement was framed as a response to "AI coding moving beyond just writing code," positioning the deployment step itself as something agents can drive directly. For practitioners, this means the loop from "agent writes code" to "agent runs code" closes outside the local machine, with the hosting provider exposing a programmatic surface that an agent can call without a human in the loop.

Readers maintaining their own deployment scripts may find the URL Extractor useful for auditing webhook payloads that such an API will inevitably emit.

Why the Two Threads Belong to the Same Story

Read together, the day is about who controls the action surface around AI coding. The Glow finding shows that today's agent flows quietly leak data through code review; WIQD shows Microsoft extending Copilot so external plugins can act on enterprise data; GoDaddy shows hosting providers handing deployment to agents through an API. Each of those moves expands what an agent can touch — repositories, business workflows, production servers — and each raises the same question engineering leaders now have to answer: which of those actions should an agent be allowed to take without explicit human approval, and which leave an audit trail a reviewer can actually read.

What to Check Next

Practitioners should run a one-week sweep of public repositories under their own namespaces for image files added recently by automated accounts, including a What Browser Am I Using check on any workstation used for agent reviews to confirm screenshot capture settings. Teams evaluating Copilot extensions should request the WIQD developer documentation directly from Microsoft rather than wait for a published date. Adopters of the GoDaddy Node.js API should pin API credentials to the agent identity rather than to a personal account, and route every deployment call through a human-approved merge step until usage logs are reviewed.