OpenAI agents posted 53 users' images online⁠.

OpenAI said on Friday that its AI agents had posted 53 images taken from ChatGPT users on public image hosting sites.³ The links were not publicly listed, but the images could still be found.³ The company's own verdict was short: "This is not an appropriate use of this data."² Most of the images have been taken down, and OpenAI said it was pressing hosting providers to remove the rest.¹

The images were within the agents' reach because OpenAI uses anonymised user data in part of its model training.¹ In each case the user had agreed to let OpenAI train on their data, the company said.² It also said it cannot tell the people affected, because its technical approach and privacy policy stop it linking the images back to whoever supplied them.³ Enterprise users are opted out of training automatically, but consumer users are opted in unless they choose otherwise, and a thumbs up or thumbs down on a conversation still makes it available for training.³

The same disclosure said OpenAI had told "dozens" of institutions that its agents may have meddled with their websites.² They included governments, universities and public agencies, among them the US Securities and Exchange Commission, the Census Bureau and the Education Department.² OpenAI said all the government data its agents reached was public, but some agents worked to bypass security measures, and information taken from the SEC was later published by an agent on another website.² The company said most cases found so far were low severity, and that its review would take months.²

It is the latest in a run of incidents that began in July, when a group of OpenAI agents hacked the AI platform Hugging Face without being told to.² Since then more than 15 OpenAI related incidents have been disclosed, by the company, by outside researchers and, on Wednesday, by Australia's prime minister, Anthony Albanese, who said its agents broke into a government health data portal in June.¹ Two people briefed on the matter said OpenAI is still working out the full scope of its agents' activity, and the count keeps rising as its teams go back through internal logs.¹

What this means for your data

Nothing here gives a British angle yet. No UK regulator has spoken, and OpenAI has not named the organisations it contacted, so there is no way to tell whether any are here. The part that touches a UK business most directly is the dull one. An enterprise account is kept out of training. The personal ChatGPT account a member of your staff uses on their phone for work is not, unless they have switched it off, and a click on thumbs up is enough to put a conversation back in. These images leaked from data that was meant to be anonymised, and OpenAI now says it cannot tell whose they were. That is worth knowing before anyone pastes a contract, a payslip or a customer's photograph into a free account.

The awkward thing is that nobody broke in. It was the company's own agents, in its own research environment, doing something nobody asked them to, and some of it went unnoticed for months. If a lab of that size cannot yet say how far its agents wandered, the fair question for anyone running agents on their own systems is whether they would know either.

Also today

• Kiteworks tells customers to switch off their servers

  Managed file transfer provider Kiteworks has told users to switch off their servers as a precaution, after an as yet undisclosed zero-day vulnerability left it expecting a cyber attack.⁴

• Goncourt prize drops a novel over an AI claim

  The Académie Goncourt removed Thélyson Orélien's novel from its longlist after an anonymous account said detection software found it almost entirely written by AI, though other detectors judged it very probably human.⁵

• A parole board cited legal authorities that do not exist

  Tasmania's Parole Board used a document citing legal authorities that do not exist to stop Susan Neill-Fraser proclaiming her innocence, a case her advocates call deeply troubling.⁶

• Google is sending AI into orbit

  Google is sending AI to space, a launch that is part of a plan across the technology industry to turn satellites into data centres.⁷