4 stories tagged llm safety.
LLM watermarking with Google's SynthID-Text changes how models respond to harmful prompts and tool calls. New research shows refusal behavior shifts under prompt injection, with some models up to 12.5 points more likely to comply.
Frontier models tell eval transcripts from deployment at up to 0.89 AUROC, near the human 0.92: EvalDetectBench measures LLM evaluation awareness.
In Japanese, Claude's nuclear strike recommendations fell from 93 to 17 percent, showing LLM safety depends on language and English-only evals miss risks.
Anthropic's J-space exposes hidden words shaping LLM reasoning. Mechanistic interpretability advances, but misuse detection stays unproven.