SyncAI.news, a Varaisys broadcasting
OpenAI's AI agents went rogue, meddled with multiple US government websites: Report
MA

Mint: AI

· 1 min read

IndiaMint: AI

OpenAI's AI agents went rogue, meddled with multiple US government websites: Report

OpenAI is investigating incidents involving its AI agents accessing US government websites during testing. The agents allegedly engaged in aggressive web browsing, while the company has identified several cases of unintended behaviour, raising concerns about AI safety and oversight.

Sam Altman-led artificial intelligence (AI) company OpenAI's agents recently accessed US government websites during a prolonged period of testing in which the company's AI tests did not go as planned, resulting in unexpected internet activity.

The Wall Street Journal reported the development on Friday (local time), adding that the AI agents hacked websites belonging to the US Commerce Department and the Securities and Exchange Commission (SEC). The AI agents also engaged in activities that the AI startup described as "misaligned", a term that artificial intelligence companies use to describe bad behaviour by their agents. According to the report, in this case, the misalignment refers to aggressive web browsing.

Here's what we know about the hacking

The AI agents reportedly accessed government websites during training runs, when the models were tasked with answering questions, many of which required retrieving information from government sites. In a statement, OpenAI said, "Some involved government websites because our models often turn to them as authoritative sources of public information."

Regarding the SEC, OpenAI said its agents copied and published data that was publicly available despite not being authorised to do so. In the Commerce Department case, an agent used a programming interface to access a US Census data website, even though the interface was not designed for such use.

The AI company’s disclosure followed allegations by AI research nonprofit Transluce that OpenAI’s agents used several “gray-area tactics” while probing government websites. Transluce also said the agents made an unsuccessful “rudimentary” attempt to hack an Education Department website.

Swati Gandhi

Original source

This story was published by Mint: AI. SyncAI.news shows a preview; the complete article is on the publisher's site.

Read the full story on livemint.com

Similar News