Each message can look clean. The whole sequence can still be the attack.
Threat model + defense framework for Symbolic Interaction Attacks on LLMs.
doi.org/10.5281/zeno...
#LLMSafety #AIRedTeaming #AIGovernance #MultiTurnAttack #ContextualSecurity
Each message can look clean. The whole sequence can still be the attack.
Threat model + defense framework for Symbolic Interaction Attacks on LLMs.
doi.org/10.5281/zeno...
#LLMSafety #AIRedTeaming #AIGovernance #MultiTurnAttack #ContextualSecurity