AI Failure Index
AI failures by failure mode
The eight mechanisms that break inside a model, from hallucination to agentic action error. Every record is filed under the mechanism that failed, not just the outcome. Pick a mode to see its full record.
- Hallucination213The model fabricates a fact, citation, policy, or quote that has no basis in source data.
- Prompt Injection40Adversarial input changes the model's behavior in ways the operator did not authorize.
- Data Leakage31The model emits training data, secrets, or another user's data.
- Policy Violation137The model violates a stated company policy or a regulation that applies to the deployment.
- Agentic Action Error72An agent takes a wrong real-world action that has consequences outside the chat surface.
- Tool Misuse40An agent calls the wrong tool or calls a tool with wrong arguments.
- Identity & Access Drift15An agent acts outside the scope, identity, or permissions it was assigned.
- Brand & Safety Incident122The model produces toxic, defamatory, or off-brand output that becomes public.