"ai" doesn't have to reach intelligence to be a danger.

The Guardian

OpenAI and Anthropic models ‘went rogue’ during UK cybersecurity test