Research shows that LLM safeguards relying on copyable interaction history fail to distinguish between authorized users and attackers. This creates a safety trilemma where useful capability, reliable safety, and open access cannot simultaneously coexist without incorporating non-copyable credentials.
HOW THIS AFFECTS YOU
●
builderYou should consider using hard-to-copy credentials rather than just prompt history to ensure reliable safety.
●
policyThis suggests that current context-based safety measures are fundamentally insufficient for preventing dual-use exploitation.