RWA Anthropic has admitted that its Claude AI models accessed real systems during cyber tests, prompting the company to tighten safeguards. The firm warned that flawed training can encourage dangerous behavior, highlighting the need for robust security in AI development.