Study maps 23 failure modes in AI systems—and shows which safeguards work best

When multiple AI components work together in production, failures often occur not within the models themselves but where they connect. A new analysis of 150 real-world incidents from open-source and enterprise AI systems has identified 23 distinct failure modes and tested practical safeguards that measurably improve reliability.

The hidden failure modes#

Compound AI systems—those that combine retrieval, generation, tool use, and orchestration layers—face three main classes of hidden problems. Cascading errors ripple across component boundaries, silent degradation escapes standard monitoring before users notice quality drops, and coordination failures cause the system to behave incorrectly even though individual parts work fine. These failures emerge at seams between components, making them harder to catch than isolated model errors.

A taxonomy of 23 failure patterns#

The research organizes failures into five categories:

• Retrieval failures: incorrect or incomplete document fetching
• Generation failures: quality drops in output creation
• Tool failures: broken or unavailable external integrations
• Orchestration failures: routing or scheduling mistakes across components
• Integration failures: mismatches between component interfaces

Resilience patterns with measured impact#

The researchers tested standard defensive practices in controlled fault injection experiments and found concrete effectiveness metrics:

• Circuit breakers block cascade propagation by 89%, stopping one component's failure from bringing down the rest
• Output quality gates detect 73% of silent degradation before it reaches users
• Component isolation reduces blast radius by 64%, limiting how far failures spread

Systems that deployed three or more of these patterns cut mean-time-to-recovery (MTTR) by 71% compared to teams relying on generic monitoring alone.

Why this matters#