Taxonomy

AI Agent Failure Taxonomy

97 documented failures, classified by what broke — and, for 97 of them, by why it happened.

These are two independent axes over the same incidents, not one flat list of categories. All 97 incidents are classified by what broke — 8 of them as systemic findings, counted at the foot of this page rather than among the failure modes. 97 carry a root cause and are classified below.

What broke

The observable failure. Excludes systemic findings, which are counted separately below.

Categories with too few incidents for their own page are listed but not linked.

Why it happened

The underlying cause. Better distributed than failure modes, and the more useful axis if you are deciding what to guard against.

Categories with too few incidents for their own page, or excluded as a catch-all, are listed but not linked.

Where the axes cross

The same cause produces different failures, and the same failure arrives via different causes. The strongest pairings:

  • Tool MisuseDestructive Action16
  • Scope MisunderstandingDestructive Action11
  • Tool MisuseSecurity Vulnerability9
  • OtherSecurity Vulnerability7

Systemic findings

8 entries carry the catch-all classification rather than a named failure mode, and are counted here rather than among the categories above. 5 of them come from benchmarks and studies rather than a single reported incident.

Browse systemic findings →

A note on one word

Logic error appears on both axes and means two different things: as a failure mode, the agent produced incorrect logic; as a root cause, faulty reasoning led to some other visible failure. The incident data is written by an automated routine and renaming it here would misrepresent the source, so the collision is stated rather than hidden.