AI Incident Severity Classifier

Classify AI safety incidents by severity and required response timeline

Ad placeholder (leaderboard)

AI incident severity classifier

When something goes wrong with an AI system — a harmful output at scale, a prompt that leaked data, a model returning another user’s information — the first job is triage: how bad is this, who needs to be paged, and how fast must we act? This classifier turns a short description of the incident into a P0–P4 severity level with the corresponding response timeline and playbook, so your first response is fast and consistent instead of ad hoc.

How it works

You describe the incident type (harmful/unsafe output, data exposure, security breach, availability/outage, or bias/discrimination), the number of affected users, and what kind of data was exposed, plus flags for regulatory-breach signals and ongoing/active status. The tool scores these factors — data sensitivity and breach indicators weigh heaviest, then blast radius, then reputational risk — and maps the total to a severity level. Each level comes with a target response window and a checklist: who to escalate to, containment, whether GDPR’s 72-hour breach-notification clock has started, and post-incident review.

Notes and use

Severity can escalate as facts emerge, so re-run the classifier when the blast radius becomes clearer — an incident that looked like P3 often becomes P1 once you discover the data was sensitive or the exposure was wider than thought. This is a triage aid, not a substitute for your incident policy: legal counsel and your DPO own the final call on regulatory notification, and your runbook owns the named on-call rotation. Use it to get the right people engaged on the right clock from minute one.

Ad placeholder (rectangle)