Cascading Failure

Definition

A cascading failure occurs when an initial malfunction or error in one part of an interconnected system propagates through dependencies to cause sequential failures in other components. In AI systems, cascading failures take on distinctive characteristics: hallucinated outputs from one model can become corrupted inputs for downstream models, erroneous decisions by one autonomous agent can trigger inappropriate responses from connected agents, and failures in shared AI infrastructure can simultaneously affect all dependent services. The speed of automated systems means that cascading failures can propagate faster than human operators can detect and intervene, and the complexity of AI system interactions makes failure chains difficult to predict, trace, or contain once they begin.

How It Relates to AI Threats

Cascading failure is a threat pattern spanning the Agentic and Autonomous AI Threats and Systemic and Catastrophic Threats domains. Within the agentic domain, cascading hallucinations occur when one agent’s erroneous output is consumed and amplified by other agents in a multi-agent pipeline, compounding errors at each stage. Within the systemic domain, infrastructure dependency collapse addresses scenarios where critical systems — financial markets, power grids, communications networks — share dependence on common AI components whose failure propagates across sectors. The increasing integration of AI into infrastructure systems means that AI-specific failure modes such as hallucination and model degradation can now trigger cascading effects in physical and economic systems.

Why It Occurs

• Tight coupling between AI components means that the output of one model directly feeds the input of another without independent verification
• Shared model dependencies create single points of failure: when a widely used foundation model degrades, all downstream applications are affected
• Automated systems propagate errors at machine speed, outpacing human monitoring and intervention capabilities
• Complex dependency chains in AI pipelines are often poorly documented, making it difficult to identify and isolate failure propagation paths
• Redundancy and fallback mechanisms designed for traditional software failures may not address AI-specific failure modes such as confident hallucination

Real-World Context

While no specific incidents in the TopAIThreats taxonomy currently document AI-driven cascading failures, historical analogues in other domains — including the 2003 Northeast blackout and flash crashes in financial markets — illustrate how interconnected automated systems can fail catastrophically. As AI is integrated into critical infrastructure, financial systems, and supply chain management, the potential for AI-specific cascading failures grows. Regulatory attention is increasing through the EU AI Act’s provisions on systemic risk from general-purpose AI models and through critical infrastructure protection frameworks that are being updated to address AI dependencies.