Emerging signals, last 90 days
Categories whose share of dated incidents in the last 90 days is significantly higher than in the previous 365 days (log share-ratio z ≥ 2, at least 3 incidents). Share, not count, so overall growth in AI coverage does not flag everything.
Highest-severity recent incidents
Severity 4–5 in the window, by severity then community attention.
All tested categories
Every category tested, sorted by z. Flagged rows are marked ▲.
Trends
Counts reflect what gets reported and curated, not true incidence. Read shares more than raw counts. The current year is partial.
Incidents per year by MIT risk domain
Canonical incidents after cross-source deduplication.
Share of incidents involving a frontier-LLM vendor
OpenAI, Google, Anthropic, Meta, xAI, Microsoft, DeepSeek, Mistral. Band = 95% Wilson CI.
Failure-mode share by year
Percent of that year's incidents tagged with each failure mode (multi-label, so columns do not sum to 100). Darker = larger share.
Model family, last two years
Sector, last two years
Sector of deployment as classified.
Mean severity by year
1 = near-miss … 5 = loss of life / systemic. Bars = 95% CI.
Harm type by year
Multi-label counts.
Does the classifier agree with humans?
Gemini labels were produced blind to the reference labels. Cohen's κ with 95% bootstrap CIs. The CSET row compares against human–human agreement between CSET annotators on the same incidents, the realistic ceiling.