Tag

#llm safety

2 stories tagged llm safety.

Stacked bar chart showing LLM nuclear strike recommendation rates for Claude Sonnet 4.6. Unnecessary strike scenario: 40% in English drops to 0% in Japanese. Contested strike scenario: 93% in English drops to 17% in Japanese. No-strike recommendations fill the remaining proportion to 100%.

Language switches warp LLM safety in high-stakes calls

LLM safety is language-dependent. Japanese prompts cut nuclear strike recommendations from 93% to 17% in Claude, revealing English-only evaluation misses real risks.

Lars Cornelissen · Aug 15, 2026

Slope chart showing interpretability visibility before and after the J-space discovery. Five behavior categories improve: task progress from 20 to 70, recognition from 10 to 60, commentary from 5 to 80, bias detection from 15 to 55, and cheating signals from 0 to 70. Illustrative values.

Research

Anthropic J-space reveals LLM thoughts, but not fully

Anthropic's J-space exposes hidden words shaping LLM reasoning. Mechanistic interpretability advances, but misuse detection stays unproven.

Lars Cornelissen · Jul 14, 2026