Live2026-09-21 06:00 UTC
HomeAI Risk IndexIncidents

AI Incident Record

2 incidents since 18 September 2026 · permanent

The running record of AI going wrong, as the world’s newsrooms reported it. An incident is any AI headline our engine weighed at 55 or more — harm done, attempted or narrowly avoided. Each one is numbered, dated and kept for good.

Where the record begins

Incident #1 · 18 September 2026 · BBC Business
Uncontrolled AI could lead to 'silicon species' rivalling humans, warns Microsoft 59

Recorded by engine v4.4 on the day it was published. Everything after it is numbered in order.

2026 →

2

incidents recorded

Incidents by kind of risk

all time

Misuse & cyber

1

Capability jumps

2

Regulation & control

1

Military AI

0

Safety incidents

1

Jobs & economy

0

Chips & power

0

2026 so far

newest first · full year page

September 2026

2 incidents

How this record works

Append-only. An incident enters the moment its headline is read and scored. Nothing is edited or deleted afterwards; corrections, if ever needed, are added as new lines. That is what makes it citable in twenty years.

The threshold. 55 on the engine’s 0–100 AI-risk weight: the level where a story is about harm, not a product launch. The weight comes from our own keyword tiers, sharpened by a language model when available.

What counts. A deepfake that moved money. A model that resisted a shutdown test. A drone that chose its own target. A leak of model weights. A cyberattack run by an AI agent. The common thread is harm, real or attempted, with AI as the actor or the tool.

What does not count. Product launches, funding rounds, benchmark bragging and opinion pieces about the future. They feed the capability and economy sub-indexes, not this record. Regulation stories count only when a harm is the subject.

Why it matters. Every AI-safety debate reaches for examples. This is the list, with dates and sources, kept by a system that had no stake in the argument.

Use it freely. ai_incidents.json and CSV, CC BY 4.0, attribution World Tension Meter. Method and versions: AI Risk methodology.