Evidence
Examples of AI Failure Modes
AI failure modes are easier to understand through examples. This page should show concrete patterns that appear in LLM and agent systems, with short explanations of why each one matters and how it can be detected.
In scope
Concrete failure patterns
Agent calls the wrong API
Tool misuse — the agent chooses or invokes a tool that doesn't match the task intent.
LLM fabricates a policy
Hallucination — the model produces an authoritative-sounding answer with no source.
RAG system retrieves stale content
Retrieval failure — the answer is grounded in outdated or irrelevant sources.
Agent loops through repeated tool calls
Infinite loop and cost runaway — the workflow makes no progress while spending tokens and time.
Model upgrade breaks structured output
Model regression and schema violation — outputs that used to validate now fail downstream parsers.
Prompt injection changes agent behavior
Prompt injection and unsafe escalation — external content steers the agent into actions it shouldn't take.
Where FailureModes.ai fits
Each pattern maps to a page in the AI Failure Modes Library, with symptoms, detection signals, severity, eval strategy, runtime monitoring, and mitigations.