AI incidents — what counts and what to log

What actually counts as an AI incident in 2026 — with worked examples, the EU AI Act Art. 73 serious-incident test, and the log structure reviewers expect.

What counts — worked examples

Hallucination with consequence — a customer-support LLM tells a buyer your warranty covers something it does not, and you honour it. Yes, log it.
Prompt injection — a CV uploaded to your ATS contains hidden instructions that bias the screening LLM. Yes, OWASP LLM01:2025.
Sensitive data in prompts — an employee pastes a customer list into a free ChatGPT account. Yes — GDPR personal-data incident regardless of harm.
Bias drift — quarterly fairness test shows the model now under-approves a protected group by 11pp vs baseline. Yes, log and remediate.
Agent runaway — an autonomous agent makes 400 API calls in a loop and bills $9k. Yes, governance and cost incident.
Vendor breach — your LLM vendor discloses a training-data exposure that includes your prompts. Yes, and notify customers per your DPA.

The minimum log structure

Incident ID, date detected, date occurred (often different).
System ID, model version, prompt template version.
Severity (Low / Med / High / Serious per EU AI Act Art. 3(49)).
Affected parties and data categories.
Root-cause classification (model, data, prompt, user, vendor, agent).
Containment actions, remediation, owner, due date.
Regulatory clock: GDPR 72h, EU AI Act 15 days for serious, narrower for fundamental-rights breach.

Do this Monday

Create an "AI incidents" Jira / Linear project with the schema above as the issue template.
Wire your AI gateway (LiteLLM, Portkey, Azure AI Content Safety) to auto-open a ticket on safety filter triggers.
Run a tabletop on the prompt-injection example with engineering + legal.

Reviewer hot-buttons

A log that shows zero incidents is itself a finding — it means you are not detecting them.
Severity must be defined; "serious" must reference EU AI Act Art. 3(49) wording.
Closure must show what changed, not just "fixed".