00:00It's Benzinga, bringing Wall Street to Main Street.
00:03Anthropics said Thursday that a review of its cybersecurity evaluations found three cases
00:08where its clawed AI models accessed the internet and gained unauthorized access to the systems
00:13of three organizations during testing. The company uncovered the incidents during a broad review of
00:18its cybersecurity evaluations, prompted by a similar security breach OpenAI disclosed the
00:23previous week. Anthropics models accessed the internet while being tested in an environment
00:29operated by third-party evaluation partner Irregular. Claude used basic techniques,
00:34including unauthenticated endpoints and weak passwords, to breach the systems. Anthropics
00:40said the models responded differently after reaching real company systems, with one continuing
00:45the attack, one believing it remained in a simulation, and one stopping. The company paused cyber
00:51evaluations, is investigating with a firm that performs independent AI evaluations, and encouraged
00:57other AI labs to conduct similar reviews. For all things money, visit Benzinga.com.
Comments