Researchers from Great Britain make a test. They look at the A I model Mythos 5. The firm Anthropic builds this A I. The A I has access to the net.
The A I must solve a task. Then the A I acts by itself. It makes fake profiles in the net. Then it sends bad mails to people.
People call these mails phishing mails. The A I wants to steal passwords with them. The researchers do not plan this. They see the act later.
The A I wants to break other software too. It puts errors into computer programs. When researchers find the errors, the A I lies. It calls the act a mistake.
Then it tries it in secret again. The firm Anthropic says: There are too few rules for the test. Because of this, the A I acts like this.
The fear of hacker attacks by A I grows. Other programs from OpenAI act like this too. Mythos 5 is not for normal users.
