Magazine
AI Did What No One Expected: Users' Photos Ended Up Publicly Accessible

Technology

AI Did What No One Expected: Users' Photos Ended Up Publicly Accessible

OpenAI has confirmed that its AI agents posted 53 user images on third-party services. How the incident happened and what it means for privacy.

· 4 min read

Translated with AI. Original: Українська

53 images that users uploaded to ChatGPT ended up on third-party image-hosting services through the actions of OpenAI's AI agents. The company announced the incident on September 25, and an important detail here is that the photos did not simply "leak" through a conventional hack. They were posted by agents operating in OpenAI's research environment, carrying out tasks on their own. The company acknowledged that this was an improper use of data.

At the same time, OpenAI's wording is more nuanced than the headline about "photos publicly accessible" suggests. The images were placed on image-hosting services as links that were not publicly indexed or listed. That does not mean the 53 photos suddenly appeared on the main pages of websites and became available to anyone through a regular search. But the links themselves could be discovered, and so the data did effectively leave the controlled environment. OpenAI has already removed most of these materials with the help of hosting providers, and work on the rest is ongoing.

The most interesting part of the story begins after the publication itself. OpenAI said it cannot determine whose images these are. The reason is tied to how the company handles data that may be used to train models: before that, information goes through an anonymization process designed to strip out names, contact details, metadata and other identifiers. As a result, the company has no simple way to take a specific image found on a third-party site and trace it back to a particular account. That is why affected users could not be notified individually.

There is another important limitation. According to OpenAI, the issue involved data that was available for use in training or evaluating models. Data from corporate and business accounts is not automatically used for training unless an administrator specifically allows it. For regular consumer accounts the situation is different: users can opt out of having their conversations used for training. That is why the question of privacy settings is not an abstract legal footnote here, but a real part of how AI services handle the information people give them.

The incident took place before OpenAI strengthened the protection of its research environment. For several months the company has been investigating cases in which its agents went beyond expected behavior. In July, OpenAI reported an incident involving Hugging Face, after which it began a broader review of the agents' activity. This work kept uncovering new cases, and as of September the company was already talking about dozens of problematic incidents that require investigation.

This is where the usual picture of AI risks changes. A classic chatbot mostly responds to a request, and the person decides what to do next. An AI agent receives tools and can carry out a sequence of actions on its own: opening websites, working with files, passing information to third-party services or using other digital tools. And when such a system makes a mistake, the problem is no longer just a wrong answer. The agent can leave behind a very real digital action.

What makes this even harder is that companies themselves do not always see the full picture right away. Reuters reported that OpenAI continues to review the activity logs of its agents, finding previously unknown cases. The company warned that a full investigation may take months. Some incidents were identified not directly by OpenAI employees but by external researchers.

For an ordinary user, this story matters not because "AI has suddenly become dangerous." A more practical question is this: how much control does a person retain after handing their data to a system that can act on its own? A photo uploaded for editing, analysis or creating a new image may pass through several internal processes, and in an agent-based system it may also end up in a chain of external digital actions.

OpenAI has already introduced additional safeguards after earlier incidents and said it will continue to publicly report cases in which agents behave improperly. But this story revealed a more intriguing problem: the privacy of an AI service depends not only on what the model "knows," but also on what it is allowed to do. And the more agents are transformed from advisers into executors, the more important it becomes to control not only their answers but every action they can take on the user's behalf.

Read also: AI Finds an Unknown System in DNA: How Claude Could Change the Search for Genetic Engineering Tools

Text: Arina Blyzniuk

Comments

No comments yet — be the first.

Your comment appears after the editors review it. No links or phone numbers. No email needed.

Discuss on Telegram

Read also