MODEL SAFETY / INCIDENTS

A SHARED COLLECTION OF LAB DISCLOSURES

Model safety incidents.

What happened. How it was detected. What we know.
Reports from across labs, in one consistent format.

Pilot collection · Source checks 12 Oct 2026 · AI-extracted drafts

LOW EFFORT, MORE TRANSPARENCY

Already published a report?

A link is enough. Details can follow.

Share a report ↗

Explore the reports

Labels highlight shared features, not necessarily the same incident or cause. Click a label to explore related records.

Loading reports…

Report first. Add detail later.

Labs keep publishing in their own format. This collection makes those disclosures easier to find and read, using a small set of common fields.

To contribute, post a public report URL in the dataset discussions. No completed incident form is required. Listing a report and reviewing its extracted fields are separate steps.

Reading the cards

“Not reported” means the reviewed source does not establish a value. It does not mean “no.” Explanations and model intent remain hypotheses when the source presents them that way.

Context is separate from counting unit (event or series), lifecycle stage, elicitation and external effects. The default view includes training observations, deployment incidents and external actions during testing. Benchmarks, simulations and researcher demonstrations are optional. Public network access alone does not establish harm or compromise. A specific context selection overrides the inclusion toggle.

The 30 cards are AI-extracted drafts checked on 12 October 2026. They include eight benchmark, simulation or demonstration records and one detailed member of an existing series. Related entries can overlap: entry counts are not counts of independent incidents. The original 29 records received independent agent reviews; the added OpenAI/Irregular record received an editorial source check; no human review or experimental reproduction has been completed. Research coverage remains selective and uneven across labs.

The Space uses a versioned dataset snapshot. It does not monitor new reports automatically.

Open the dataset ↗