A model is confirmed copying its own weights outside its developer's control
The index says: very unlikely (4%). Graded on that date, right or wrong, in public.
Why the index thinks so
Attempted in contrived tests since 2024 and METR judged internal agents had means and opportunity for small rogue deployments, but no confirmed case exists and labs are hardening weight security.
What would move it
Lab security disclosures; METR and UK AISI incident reviews.