Open data

Use the record in your own work

The full record of 35 incidents is available as one file, for researchers and developers. It is rebuilt every time the site is published.

License

The data is published under CC BY 4.0. You may share and adapt it if you credit Incidents.si and link to the entry or to this page. The linked source articles belong to their publishers and are not part of the license.

Version

Schema version 1.0. The JSON file states the version and the date it was generated. Field names do not change within a major version. Incident IDs never change.

What to keep in mind

  • The number of entries is not a failure rate for any system or organization.
  • For research-demo and near-miss entries, severity is potential. The field severity_is_potential marks them.
  • The fields patterns and takeaway_* are our editorial conclusions. Everything else is drawn from the cited sources.

Fields

FieldDescription
idPermanent identifier, for example INC-2026-0001. Never changes.
urlAddress of the entry on incidents.si.
titleWhat happened, without value judgments.
summaryOne-paragraph summary.
date_occurredWhen the incident happened, if the source states it. ISO date or empty.
date_disclosedWhen the incident became public. ISO date.
date_publishedWhen the entry was published here. ISO date.
date_updatedLast update of the entry. ISO date or empty.
categoryOne of: agent-action, prompt-injection, data-leak, outage, model-behavior, supply-chain.
severitySEV1 (critical) to SEV4 (low). See the methodology for thresholds.
severity_is_potentialtrue for research-demo and near-miss entries, where severity rates the effect that would have occurred.
evidenceOne of: confirmed, reported, research-demo, near-miss.
statusEditorial status: active, disputed or retracted.
systemThe AI system or type of agent involved.
vendorThe organization, as the source names it. May be empty.
impactThe measurable effect.
root_causeThe cause, if the source states it. May be empty.
patternsFailure patterns assigned by Incidents.si (editorial). List.
takeaway_failureEditorial: the failure pattern in one sentence.
takeaway_controlEditorial: a preventive control.
takeaway_lessonEditorial: the transferable lesson.
tagsFree keywords. List.
sourcesList of sources with title, url, type (primary or secondary) and date.

In the CSV file, lists are joined with " | " and sources are given as "type: url".

Vulnerability watch

The vulnerability watch is a separate dataset under the same license: vulnerabilities.json. Each item has the CVE or GHSA id, its kind (found-by-ai or in-ai-tool), the software, a summary, the disclosure date, the CVSS score and vector as the CVE record gives them, whether it is exploited, who found it, the fixed version and the sources.