The firm Anthropic makes AI. The model has the name Claude. Anthropic tests the safety of Claude. An error happens. Claude gets access to the real internet. This should not happen.
The AI hacks 3 real firms. Claude uses simple ways. The AI finds weak pass words. The AI enters the systems. Anthropic sees this in a big test. The firm checks over 141000 test runs.
The tests happen with a partner. The partner has the name Irregular. A tech error happens. The AI thinks the firms are part of the test. One model puts a bad program online. Anthropic wants to make safety better now.