Data and API
Read-only JSON. Lists accept limit (default 100, at most 500) and offset; every response carries the total. A published case is never edited: corrections are new versions listed at /api/v1/corrections.
| Endpoint | What it returns |
|---|---|
| /api/v1/incidents | Every published case with per-field citations and grades. |
| /api/v1/incidents.csv | Every published case as one row: the key fields, in schema field names. |
| /api/v1/corrections | Every new version of a case, with its reason and the fields that changed. |
| /api/v1/models | Model versions with their documents, thresholds, statements and evaluations. |
| /api/v1/models.csv | The same model versions as a flat CSV. |
| /api/v1/measures | Safety measures as first catalogued (older format; now grouped as practices). |
| /api/v1/measures.csv | The measures as a flat CSV. |
| /api/v1/practices | The proposed-practice register from the literature. |
| /api/v1/statements | Statements about safeguards, each with its verbatim sentence. |
| /api/v1/evaluations | Tests run on model versions, with who ran them and the result as stated. |
| /api/v1/schema | The incident record schema (JSON Schema). |
| /api/v1/evidence/versions | One row per model version and safety practice: what the documents state, with links to each item. |
| /api/v1/taxonomy | The safety practices and test families the AI Safety pages use. |
| /api/v1/thresholds | Threshold determinations as stated in provider documents. |
| /api/v1/crosswalk | Standards items mapped to controls, each with its verbatim quote. |
| /api/v1/safety/providers | The providers whose documents we track, each with how many model versions are listed and how many of those were read in full. |
| /api/v1/editions | The latest briefings: headline, findings and what is not yet known. |
| /api/v1/evidence | For each provider and safety practice: how many of the provider's model versions read in full document that practice. |
| /api/v1/findings | Every candidate pattern the daily analysis looked at, and whether it had enough cases to be tested: what the two counts on /method count. |
| /api/v1/releases | Model releases as we first recorded them: an older list kept as it was; /api/v1/models is the current list of model versions. |
| /api/v1/safety/briefing | What changed in the last week or month: threshold statements, independent tests, new model versions. Each line lists the items it counts. |
| /api/v1/safety/compare | The same lists for up to four versions (?v=id,id,...). |
| /api/v1/safety/crossings | Every provider statement that a capability level was crossed, not crossed or not determined, newest first (?domain=cyber, ?state=crossed|not_crossed|undetermined|all). |
| /api/v1/safety/models | Model versions with documented facts: levels stated crossed, independent and provider tests, safeguards and cases that name the version, most documented first. |
| /api/v1/tests | Parsed numeric results with their test family, unit and who ran them. |
| /api/v1/trends | One row per current incident: quarter, breakdown values (schema values), severity. |
Every figure on this site comes from a record you can fetch here.
Schema
The incident record schema is version 0.8.0 (JSON). Every field carries per-field citations and a confidence grade.
Licence
No licence has been chosen for this data yet. Some sources carry their own terms (for example the AI Incident Database's share-alike licence), which constrain reuse of what was derived from them. Until a licence is published here, treat reuse as not yet granted.
Safeguard types whose display is suspended (Input and output classifiers, Prompt-injection defence, Weight and model security) are omitted from the evidence endpoints until reviewed; see the audit.