Incident monitoring

AI Incident Register

Documented cases where an AI system acted outside the control of the party operating it, classified by how far the artefact involved can be traced to a verifiable source.

Build

 

Geographic distribution Select a country to filter the register

Artefact provenance

Human control over the action

Most affected countries in the selected period

New cases per month stacked by provenance class, dashed line: cumulative

Record quality what this register does and does not establish, across all entries

Register

Method

Provenance classes

Certified
The artefact traces back to a verifiable publication: a known registry, plus at least one element that can be checked afterwards: model card, declared licence, signed release, pinned version, vendor contract.
Suspicious
At least one signal of tampering or masked origin: unsigned weights, imitative naming, serialised payload, copy without lineage, deployment outside approved channels.
Not established
No source states which system was involved, or the registry is known with nothing verifiable attached.

A tampering signal always overrides the registry. A model published on Hugging Face that impersonates another publisher is suspicious, not certified: the registry says where the file sits, not who made it.

Severity

  1. Inconvenience or embarrassment, no lasting consequence
  2. Contained, reversible harm
  3. Economic or rights harm at company or local scale
  4. National scale, injury, or fundamental rights of many people
  5. Death, or systemic harm to a population

What this does not measure

Coverage follows disclosure. Incidents surface where there is a free press, regulators that publish decisions and security research that discloses. Industry surveys suggest undisclosed incidents are the majority; none of them appear here. Severity is assigned by hand on a coarse scale. Provenance classification depends on how much sources say about the artefact, which is usually very little.