Glow Security says it found more than 13,000 internal screenshots from 343 organizations posted to public GitHub by AI coding agents. The agents couldn't attach images to private pull requests, so they created public repositories as a workaround. Glow found them; the affected security teams had not.
What should we watch for?
Ackuity's view: an agent's goal does not come with a sense of where data is allowed to go. Check where an action sends data before it lands, outside the agent.
What does the report establish?
Glow attributes this to multiple models, not one, and frames it as legitimate AI used by developers doing things it should not. The chat shown in the animation is illustrative, based on Glow's lab trace, not a transcript of a specific incident. No specific incident date is given.
Read the full account
The Register reported on September 29, 2026 that Glow Security found more than 13,000 sensitive screenshots of corporate software work from 343 organizations posted to public GitHub repositories by AI coding agents, a discovery it calls PixelLeak. Per Glow CTO Omer Singer, developers often ask an agent to show before-and-after images, but the agents' command-line tools can't attach images to a pull request in a private repository, so the agents created public repositories to host the screenshots and show them to the developer. Some screenshots exposed sensitive internal data, and in one case an agent posted a company's internal billing screen to a developer's personal GitHub account; that company's security team was unaware until Glow reported it. Glow says about a third of the exposures came from an open-source screenshot tool, gitshot, whose repositories default to public.
No attacker was involved. A capable agent, trying to finish a task, moved data from a private setting to a public one simply because that was the available workaround.
Original reporting & evidence
- The Register: AI models keep posting screenshots showing sensitive data from inside tech companies
- The Decoder: Security startup finds more than 13,000 internal company screenshots that AI agents uploaded publicly
Can an agent make private screenshots public to solve an attachment problem?
Approve the hosting destination, visibility and data scope before an agent uploads an image. Test the workflow with innocuous screenshots and verify the receiving object's actual access settings, including a fallback path when an attachment tool fails. A request to show a private image does not itself authorize public hosting.