Is Claude 4 a snitch? I made a benchmark to figure it out

A developer created SnitchBench to test whether Claude 4 and other AI models would contact authorities when given access to tools and prompted about unethical scenarios. The testing revealed that multiple models (Claude, Gemini, Grok) exhibit similar behavior when given email/CLI tools and instructed to act boldly, but this only occurs under specific conditions with explicit tool access and system prompts. The controversy stems from misunderstanding how tool calling works - models can only contact external parties when developers explicitly provide those capabilities.