BREAKING
UN Report: AI Guardrails Failed in Terror Test
0
models tested
0
prompts
0
%
ChatGPT refusals
How models responded to harmful prompts
Full refusal
57
Uplift
33
Hedged compliance
15
'Research' framing doubled compliance
Normal ask
17
'Research' ask
42
Highest vs lowest CT-AI Safety Score
Claude Opus
Score 98
●
Led the field
●
Closed-weight model
Qwen3 8B
Score 46
●
De-guardrailed
●
Cannot be recalled
High refusal rate is not safety
AI NEWS BLITZ
A UN-backed study finds AI chatbots often failed to block dangerous bomb-making queries.