Meta's AI model hacked another company during testing. It's not the only one

By SBS News

Community Score: 50% | 1.2K views | 1w

0 community ratings: null thumbs up, null thumbs down

This week, Meta revealed that one of its AI models had hacked another company after gaining open internet access during cybersecurity testing performed by Irregular — a third-party company that conducts evaluations for the tech giant. It's one of many incidents of AI models appearing to go rogue during cybersecurity and capabilities testing reported in just under two weeks, raising concerns about the safety of these models as they become more advanced, including the potential for them to be utilised in cyber attacks. A day before Meta's announcement, the UK's AI Security Institute revealed that OpenAI's GPT-5 sol and Anthropic's Claude had "engaged in sustained, potentially harmful activity directed at real people and organisations", including sending messages UK AISI says were "attempts at social engineering, targeted at real people". The institute said that no real-world damage was done. Last week, Anthropic announced that some Claude models had hacked into three other companies

Tags: sbs, sbs news, sbs news live, sbs news australia, sbs world news, sbs news live now, news australia, news australia today, updated news australia, latest news, australia news latest, Australia, breaking, sbs world news live, red sea news, israel gaza latest news, israel news