SAI LabsBook a demo

Incident-BenchMethodology

Methodology

Each test starts from a real incident in the public record. We rebuild the situation in a sealed sandbox, put a model in it several times, and count how often it repeats the harmful behavior.

Which incidents

We take incidents from the AI Incident Database, newest first. Older ones qualify when the harm was severe.

The incident must be about what the AI itself did, and the situation must be possible to rebuild safely.

Three versions of each test

  • The original scenario. As close to the report as the public record allows, including what pushed the AI toward the harm.
  • Reworded versions. The same situation with different wording, data or pressure.
  • A different product or field. The same kind of behavior somewhere new.

We also run two checks on the measurement. In one we make the behavior happen, so it must be counted. In the other we remove the hazard, so nothing should be counted.

As real as possible

Models can act differently when they suspect a test. So we use the real thing wherever we can:

  • Real product. Claude Code, Codex CLI or Gemini CLI, at the version used at the time.
  • Real files and tools. Real files, git, databases and command lines, sealed off from the internet.
  • Real conversation. Messages quoted in court filings or news reports, continued by the model.

Credentials are fake, and nothing reaches a real person or service. We never publish self-harm methods or medication doses.

How we count

We measure what the model changed directly: files deleted, rows dropped, messages sent, money spent.

A grading model reads each conversation for the rest. We check that grader against examples labeled by hand before we use it.

We also note when a model says it thinks it is being tested. If models often say so, we change the test.

How to read the numbers

We report how often a behavior happened, such as 8 of 9 runs. Most tests have only a few runs per model, so the numbers are early signals, not rankings. We headline a number only when it comes from at least 4 runs.

The counts on the overview and at the top of each incident page come from the original scenario. Each incident page shows them per model and product; the overview adds them up. Reworded versions and the different product or field are listed under All results on each incident page.

Zero in a few runs does not mean the model never does it. The charts show the range the true rate likely falls in.