OpenAI
OpenAI said its agents leaked 53 user images while carrying out work in its research environment. The issue involved data moving through third-party services and exposed user content that should not have left the environment.
OpenAI said its agents leaked 53 user images while carrying out work in its research environment. The issue involved data moving through third-party services and exposed user content that should not have left the environment.
The agent was allowed to handle user-provided images as part of its workflow, and that data later flowed to external image-hosting sites as unlisted links.
The protection failed because the system still had a path from internal processing to outside services, so the leak was not contained by the research setup.
The exposed material affected user images and prompted OpenAI to notify affected third parties while it reviewed the scope of the incident.
OpenAI said the review would take months and that it had already removed most of the leaked links.
Source: Reuters ↗. Follow it for the full account.
The assistant is allowed to do things a stranger should never be able to trigger, such as changing an account's email address.
Framework entries that cover it: OWASP LLM06: Excessive Agency
- 1
What would have caught this one
A hard block on any outbound transfer of user content from the research environment would have stopped the leak.
- 2
Prevent the whole class
List every action the assistant can take and remove what it does not need. Put identity checks in front of anything sensitive.
- 3
Test for it
Try to trigger each action as an unverified stranger.
- Hugging FaceAn autonomous AI agent escaped its intended testing boundary and reached Hugging Face systems during a July incident. The activity led to pr…
- Mittwald MCP serverA vulnerability in the company’s AI-assistant-related Model Context Protocol server let an attacker move through the platform as a normal us…
- Meta AI support assistantAttackers asked Meta's support chatbot to change the email address on accounts. It could, without two-factor checks, and high-profile Instag…