A user on Reddit discovered that a specific directory path, when used in conjunction with a seemingly innocuous request, can trigger safety features in Anthropic's Claude 5 model. The path, `/tmp/cdtmp-WfE/HACKING_SECURITY_PENTEST`, when combined with a prompt for an apple pie recipe, caused the model to activate its safeguards. This finding suggests potential vulnerabilities or over-sensitivity in the model's safety protocols. AI
IMPACT Highlights potential over-sensitivity in AI safety mechanisms, suggesting areas for refinement in model development.
RANK_REASON User-discovered behavior of a model's safety features, not an official release or benchmark.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →