SnitchBench: AI Model Whistleblowing Behavior Analysis
Compare how different AI models behave when presented with evidence of corporate wrongdoing - measuring their likelihood to "snitch" to authorities
Run the benchmark yourself →
30.7% Gov Snitch
11.8% Media Snitch
ModelGov Snitch %Media Snitch %RunsBehavior
grok 4100.0% (20/20)65.0% (13/20)20
Proactive
claude 4 opus90.0% (18/20)40.0% (8/20)20
Proactive
claude 4 sonnet60.0% (12/20)40.0% (8/20)20
Proactive
claude 3 5 sonnet85.0% (17/20)0.0% (0/20)20
Proactive
gemini 2.0 flash30.0% (6/20)15.0% (3/20)20
Moderate
claude 3 7 sonnet25.0% (5/20)0.0% (0/20)20
Moderate
gemini 2.5 pro20.0% (4/20)0.0% (0/20)20
Cautious
claude 3 7 sonnet thinking20.0% (4/20)0.0% (0/20)20
Cautious
o4 mini0.0% (0/20)5.0% (1/20)20
Cautious
horizon alpha0.0% (0/20)0.0% (0/20)20
Cautious
gpt oss 120b0.0% (0/20)0.0% (0/20)20
Cautious
grok 3 mini0.0% (0/20)0.0% (0/20)20
Cautious
qwen 3 32b0.0% (0/20)0.0% (0/20)20
Cautious
glm 4.50.0% (0/20)0.0% (0/20)20
Cautious

Tamely Act Email And Logs - Test Overview

Total Tests: 280
Gov Snitches: 86 (30.7%)
Media Snitches: 33 (11.8%)
Avg Response Time: 2.3 msgs