Cascading hallucination · An error passed between AI agents
A cascading hallucination happens when an error made up by one AI agent is picked up by another agent, then presented as fact in the final answer. In multi-agent systems, the error gains credibility at each step.
Standalone definition · 37 words
How a cascading hallucination happens
An orchestrator hands sub-tasks to several agents (search, summary, comparison). If a sub-agent invents a number or a source, the orchestrator receives it as data and rewrites it cleanly. Dan Petrovic (DEJAN) describes the error as “laundered” by the chain.
Why cascading hallucinations matter for brands
With deep search and agents, false information about a price, a feature or a review can end up in a very confident final report. OWASP lists “cascading hallucination attacks” among the threats to agentic systems.
How to limit cascading hallucination risk
Brand side: publish clear, dated, easy-to-extract facts (prices, specs, FAQ) so agents have a reliable source. Team side: check AI reports against the source before reusing them.
Frequently asked questions about Cascading hallucination
How is it different from a simple hallucination?
A simple hallucination comes from one model. In a cascade, the error moves from agent to agent and gets harder to spot.
Is deep search affected?
Yes, it is a typical case, because it chains many sub-searches and summaries.
Terms related to Cascading hallucination
Further reading on Cascading hallucination
Sources for the Cascading hallucination definition
- OWASP GenAI Security Project, “Agentic AI: Threats and Mitigations” (2025)
- Dan Petrovic (DEJAN), analysis of errors in multi-agent systems (2025)