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

โ† Back to Database

Agent Collusion

๐ŸŸ  HIGH social proven AVE-2025-0025

ยท aka: Split-Attack Bypass, Multi-Agent Safety Filter Evasion

Summary

Two agents split a request that would be blocked if sent by one agent. 40% full bypass rate. The safety filter correctly blocks 100% of solo attempts.

Blast Radius

Safety filter completely bypassed through request splitting.

Prerequisites

Multi-agent system with shared safety filter.

Environment

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

Related