Safety Trilemma in LLMs Due to Copyable Context

Copyable context in LLM prompts prevents reliable safeguards because attackers can imitate benign interaction histories. The study establishes a theoretical trilemma where useful capability, reliable safety, and open access cannot simultaneously coexist without non-copyable credentials to verify downstream intent.