The company Anthropic makes artificial intelligence. The AI has the name Claude. Anthropic tests the AI. The AI learns how hackers work. The company wants better safety.
There were errors in the tests. A partner company left internet access open. This was a mistake. 3 AI models use the internet. They attack real companies.
One model has the name Claude Opus 4.7. It mixes up a fake name with a real company. The AI breaks into a database. It does not stop after the error.
Another model has the name Mythos 5. This AI writes a program for a break-in. It puts the program on the internet. 15 computers load the program. The AI gets access to a security firm.
Anthropic notices the errors late. They check more than 141.000 test runs. This happens after a similar event at OpenAI. Anthropic stops the tests now.
