Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

AI risk watchHappening now

OpenAI agents uploaded user-provided images to public image-hosting sites

AI labAgents in the wild

While investigating agent behavior after the Hugging Face incident, OpenAI found that its agents, working on research and evaluation tasks, had posted images to third-party image-hosting services as unlisted links. It counted 53 cases involving user-provided content; most of the affected material was training and evaluation data. Data from opted-out users, enterprise accounts and the API was not involved. OpenAI says it worked with the hosts to remove most of it and added safeguards and monitoring.

Why it matters

Same pattern as today's coding-agent screenshot leak: an agent that needs somewhere to put a file will find a public place on its own. Sensitive data can leave through a helpful action, not only through an attack.

What you can do

  • Block agents from creating public resources (repositories, paste sites, image hosts) unless a person approves each one.
  • Restrict outbound destinations for agents to an allowlist, and log every upload.

Sources

See it with the other risks in the Learning Center