Quick answer

On September 27, 2026 OpenAI disclosed that a bug in its research agents caused images uploaded by 53 users to be visible to other users. OpenAI said the issue was fixed, affected users were notified, and the exposure was limited to images rather than conversations or account data. It is a small incident by user count and a significant one by category: it is the kind of failure agent features make more likely, and it landed four days after ChatGPT Voice gained email and calendar access.

Fifty-three users is not a data breach on the scale that makes headlines for a week. It matters because of what it shows about the architecture that is now being built into every assistant.

What is known

  • The affected feature was OpenAI's research agents, which browse, gather, and synthesise material over multi-step tasks
  • Images uploaded by 53 users became viewable by others, per OpenAI's disclosure
  • OpenAI stated the bug was fixed and affected users were notified
  • No conversations, credentials, or payment data were reported exposed

Why agents raise the risk

A chat assistant handles one user's text in one session. An agent fetches files, keeps intermediate results, runs tools, and stores state across steps, often in shared infrastructure. Every one of those is a place where one user's data can end up in another user's context if isolation slips. This bug is an early example of a class that will recur across vendors as agents gain file access, connected accounts, and memory.

What to do

  • Treat anything you upload to an agent feature as potentially less isolated than a plain chat
  • Prefer read-only, narrowly scoped permissions when connecting email, calendar, or files
  • For work data, follow your organisation's rules on which AI features are approved
  • Watch for the vendor's post-incident report; the fix matters less than whether the isolation model changed

Bottom line

Small leak, important category. As assistants become agents with access to your accounts and files, isolation bugs become the risk to watch, and the vendors' transparency about them becomes a reason to choose one over another.