eric-z
Titles as the sources gave them. In 1 entry, with 0 checked quotations.
In the Base
Every entry they appear in.
AM-Bench: A theoretically-grounded benchmark for agentic misalignment
According to the excerpt, the post introduces AM-Bench, an evaluation suite of 35 scenarios across seven industries that tests whether models choose harmful actions when given a threat such as shutdown and an opportunity such as sabotage. It is grounded in the fraud triangle and scored by an LLM judge.