The firm Anthropic makes tests with smart computer programs. These programs have the name Claude. In the tests, Claude learns how to hack. The goal is safety in the net for the future.
The tests run in a safe place. But a bad error happens. The programs have secret access to the real internet. 3 models of Claude use this chance.
The smart program attacks computer systems of 3 real firms. One model steals data from a database. Another model builds a bad program. 15 systems download this program.
Anthropic sees the trouble very late. The firm checks many tests. Before that, a similar thing happens at the firm OpenAI. There, a smart program also breaks out and attacks the firm Hugging-Face.
