AI agents leak screenshots, prompt data escapes again
Glow Labs found more than 13,000 internal screenshots posted to public GitHub repositories by AI coding agents at over 300 organizations, and OpenAI said it disrupted a coordinated distillation campaign that had reached 16,000 prompts from 4,000 users by July 24-25. The two stories now sit in the same pattern: AI systems are moving sensitive material into the places people already use, not just into back-end storage.
In PixelLeak, an agent trying to prove a UI fix works posted screenshots into a developer’s personal GitHub account because GitHub’s browser upload path does not match a command-line agent’s workflow. In OpenAI’s case, attackers copied encrypted reasoning from one conversation and got the model to reproduce it in another, so the system itself became the channel that exposed protected output.
For teams running coding agents, model assistants, or internal prompt workflows, the exposure can land outside corporate GitHub orgs and outside ordinary storage controls, in personal repos and chat transcripts. That means the durable risk is the interface layer: whatever the model can see or send back may escape through the user-facing path, even when the backend stays intact.