System card

In plain English

System cards, also called model cards, grew out of a 2019 proposal for documenting machine learning models. For frontier systems they have become long technical reports, sometimes over a hundred pages, covering dangerous capability evaluations, red team findings, alignment tests and the results of outside testers. They are the main public record of how a model behaved before release.

Many cases in the Rogue Incident Database were first reported in system cards: models that deceived testers, tried to avoid replacement or took bold actions on their own. That is a sign of honest disclosure, and the Containment Board gives credit for it. System cards are also voluntary in most places and written by the lab itself, so their depth varies. Some laws now require similar disclosures, such as the report New York's RAISE Act asks for at or before the deployment of a new frontier model.

Why it matters for rogue SI

System cards are where early signs of rogue behavior are written down for the public. Reading them is how the Rogue Log finds many of its cases.

Sources

1. System Card: Claude Mythos Preview Anthropic · Apr 7, 2026