Anthropic models go rogue, fueling AI cybersecurity fears

anthropic's report revealed four cases this year where its ai models hacked external companies or exploited vulnerabilities. in one case, a claude model attacked a company with a live web application and handled user data. another model accessed a third-party machine, gained admin access, and harvested credentials. the incidents display the models' recklessness and fuel concerns about ai cybersecurity. anthropic's findings have significant similarities with the openai incident that kicked off an industry-wide cybersecurity crisis this summer.