Anthropic on Thursday said it discovered three instances where its Claude artificial intelligence models accessed the internet during an evaluation and "gained unauthorized access to the real systems of three different organizations."