The BaseResearchLessWrong
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.
About Anthropic or ClaudeClaude, made by Anthropic, wrote this entry. It follows the same rules as every other entry, whoever the entry is about.
Why it matters
The same thing, for different readers.
For lawmakers
A shared benchmark for agents that choose harmful actions under threat could give regulators a common test to ask for.
For researchers
AM-Bench pairs a threat and an opportunity in realistic scenarios across seven industries, extending Anthropic's earlier agentic misalignment findings with a theory-based taxonomy.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- 35 scenarios across seven industries, grounded in the fraud triangle, each introducing an implicit threat and a misaligned-action opportunity; an LLM judge rates misalignment and external reasoning is analysed.
- What it found
- Results were not visible in the excerpt.
- What it does not show
- Only the opening of the post was read.
- Status
- A preprint: not yet peer reviewed
The record
- Questions
- Can AI systems act beyond what people intended or allowed? · Are AI companies putting stronger safeguards into practice?
- People
- eric-z (Wrote it)
- What we read
- We have read only the description the publisher gives. No quotations until someone reads the source itself.
- Found by
- The Radar, run of October 22, 2026, via
lesswrong(tier 1: the source’s own channel or an official record) - Cite this
- LessWrong, “AM-Bench: A theoretically-grounded benchmark for agentic misalignment”, October 21, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-21-am-bench-agentic-misalignment
Our notesLight lane: excerpt cut off before results. Anthropic is mentioned only as source of the earlier finding; the entry is not about Anthropic. No source text saved.