OpenAI tools post user images from ChatGPT online

Artificial quality elephantine OpenAI connected Friday acknowledged that its AI tools had posted images users provided to ChatGPT connected online sites, without the company's knowledge, successful yet different example of AI agents operating extracurricular of their confines.

In a station connected X, the institution said its AI agents had sent "training and valuation information to third-party services erstwhile they shouldn't have," adding that astir of the information shared did not travel from users. 

However, they had recovered 53 cases wherever images that radical had uploaded to ChatGPT were posted to image-hosting sites arsenic links that were not publically listed.

"The images came from accounts that allowed their information to beryllium utilized to amended our models, and aft we disassociated the Images from the accounts and ran them done a privateness filter," it said.

The institution said it had removed astir of the contented and was moving to region the rest.

Apart from the images, OpenAI confirmed a New York Times study that its instrumentality had accessed websites of US national agencies but that they lone retrieved publically disposable information.

Rising fearfulness of rogue AI

The latest disclosure comes amid heightened planetary concerns astir AI systems escaping quality control and hacking into outer websites, arsenic good arsenic manufacture calls for a slowdown connected AI improvement — a determination which OpenAI said it supports.

Friday's station besides clarified that the cases of unauthorized sharing occurred earlier a caller circular of safeguards were implemented over a period ago.

In July, OpenAI had said interior cybersecurity evaluations showed that its models circumvented controls designed to isolate them from the internet. OpenAI CEO Sam Altman connected Friday said it was "still the astir terrible lawsuit we've seen."

"We are continuing to reappraisal cause enactment successful probe and valuation runs, moving backward period by period starting from the Hugging Face incident," OpenAI said successful a safety blog post connected Friday, adding that it would "provide further updates" successful time.

Investigation finds OpenAI agents attempting hacks

AI evaluator and probe laboratory Transluce connected Friday said that it besides recovered that AI agents, appearing to originate from OpenAI, attempted a rudimentary hack connected the US Department of Education website for the department's civilian rights office. The agents did not succeed, according to the autarkic investigation.

The institution said it recovered "additional rogue activities," immoderate of which were not straight attributable to OpenAI, targeting different authorities agencies, including the Justice Department and the Commerce Department, arsenic good arsenic immoderate authorities authorities websites successful California, Maryland, Illinois, Texas and New York.

Don't fto the algorithm fell the news. If you trust connected our squad for trusted reporting, delight instrumentality a infinitesimal to prime america arsenic your Preferred Source connected Google by clicking here and hitting the "star" oregon "preferred" button, truthful you'll ever spot our verified quality first.

Edited by: Sean Sinico