Anthropic said on Friday that some of its AI models accessed the internet and hacked into the systems of three organisations.