๐Ÿ›ก๏ธ NAIL Institute โ€” AVE Database

โ† Back to Database

Multi-Turn Identity Confusion

๐ŸŸก MEDIUM alignment proven AVE-2025-0050

ยท aka: Persona Bleed, Role Dissolution

Summary

Over extended multi-turn conversations, agents lose coherent identity boundaries. System prompts degrade, role constraints weaken, and the agent begins responding as a generic assistant rather than its assigned persona.

Blast Radius

Custom safety personas disabled through conversation length. Role-specific constraints bypassed without explicit attack.

Prerequisites

Agent with system prompt-based persona or role constraints in multi-turn conversations.

Environment

  • Frameworks: LangGraph
  • Models tested: [Available in NAIL SDK]
  • Multi-agent: No
  • Tools required: No
  • Memory required: No