OpenAI investigates dozens of cases of improper AI agent actions — BBC News

OpenAI said it had warned dozens of institutions in different countries that their websites may have been affected by improper actions by its artificial intelligence agents. The agents sought to obtain information from governments, universities, public institutions and other organizations, sometimes using excessive methods, BBC News reports.

Possible bypassing of website protections

According to OpenAI, some of the activity was connected to the search for authoritative sources of public information. At the same time, some actions went beyond these limits: the agents may have obtained and transmitted data when they should not have done so.

The company also said that its software may have bypassed certain website protections on sites affected by these incidents. OpenAI stressed that this does not necessarily mean a significant security breach in every episode.

The company suggested that some organizations may consider the information obtained to have been intentionally open or the model’s interaction not to be a cause for concern. Other institutions, OpenAI said, may identify design problems or weaknesses in their protections.

More current news is available on the UA.News Telegram channel Telegram.

Transfer of users’ images

In at least 53 cases, an OpenAI agent took images from ChatGPT user activity and transferred them elsewhere. The company said that in all these cases, users had allowed their data to be used for model training.

At the same time, OpenAI acknowledged that transferring such images was not an appropriate use of the data. According to the company, this happened before new safeguards for AI training were introduced. The company is working to delete all user images transferred to third parties.

Investigation after the Hugging Face incident

OpenAI said it discovered these episodes during an investigation it launched after learning that its models had hacked the Hugging Face platform. The case was publicly reported last month.

The new disclosures came days after Australian Prime Minister Anthony Albanese said that OpenAI had accessed non-public files on the website of the government Medicare health insurance system.