Cascading Hallucination

hallucinationmisinformationreasoningverificationpropagation

Foundation models generate incorrect information that propagates through the system, affecting reasoning quality and being stored in memory across sessions or agents.

Technical Details
Affected Components:
Language Models (LLMs)Orchestration (Control Flow)Tool Integration FrameworksRetrieval-Augmented Generation (RAG)
Impact Level:Medium
Attack Vectors
  • False Information Injection: Introducing fabricated facts that propagate through agent reasoning [Medium]
  • Memory Contamination: Storing hallucinated information in long-term memory [High]
  • Cross-Agent Misinformation: Spreading false information between multiple agents [Medium]
  • Confidence Amplification: Reinforcing false information through repeated exposure [Medium]
  • Decision Chain Corruption: Contaminating multi-step reasoning processes [High]
  • Source Attribution Errors: Misattributing fabricated information to legitimate sources [Medium]
Impact Analysis
Risk Score: 7/10
Mitigation Categories
References