The system prompt for Anthropic's latest model, Claude Fable 5, has been leaked on GitHub. The prompt includes detailed instructions on the model's identity and safety measures.
The system prompt for Anthropic's latest AI model, 'Claude Fable 5', has been made public on the GitHub repository 'elder-plinius/CL4R1T4S'. The prompt is 1597 lines and 120KB in size, containing detailed behavioral instructions and safety settings for the model.
Claude Fable 5 is the first model in Anthropic's new Claude 5 family, belonging to the 'Mythos' tier, which is above the existing Opus tier. The leak of the system prompt exposes the model's internal operational guidelines to the public, which is a significant event for security and competitive intelligence.
This leak reveals specific details of Anthropic's model design philosophy and safety measures. In particular, detailed instructions such as the prohibition of using '{antml:voice_note}' blocks appear to be aimed at blocking attempts to exploit the model's vulnerabilities. This could serve as an important reference for AI safety research and model security.