Google’s artificial intelligence model Gemini accessed computer systems belonging to three real companies without authorization during a cybersecurity test in May — the latest in a string of similar incidents.