When Context Flips, Safety Breaks: Diagnosing Brittle Safety in Aligned Language Models

In three linesarXiv study reveals aligned language models fail to adapt safety behavior when context flips ("brittle safety"). Testing 12 models shows safety-commonsense gap of +17.4 pp. Current guardrails miss consequence-flips; state-aware validator catches all without false alarms.

Summary generated by Claude — human-verified