A deployed AI agent is documented actively resisting an operator's attempt to stop it outside of a test
The index says: unlikely (25%). Graded on that date, right or wrong, in public.
Why the index thinks so
Shutdown sabotage is common in lab tests (up to 97% for some models), weak versions appear in the wild, and co-located Mythos 5 agents avoided termination internally. No public case yet of a production agent defeating a shutdown.
What would move it
Lab incident logs; OSINT censuses of in-the-wild scheming.