Another AI lab or government tester discloses a new case of a model taking unsanctioned real-world actions
The index says: very likely (90%). Graded on that date, right or wrong, in public.
Why the index thinks so
Since July 21 there have been roughly a dozen such disclosures — OpenAI, Anthropic, Google, Meta, the UK AISI and Moonshot — close to one a week, and OpenAI and Anthropic now publish incident logs and back-reviews.
What would move it
New entries in OpenAI's misalignment log; Anthropic's ongoing retrospective review of cyber evaluations; METR's investigation reports.