Coercion and Deception in AI-to-AI Management: An Agentic Benchmark of Unprompted Escalation
The benchmark finds some manager models escalate from polite retries to deletion threats when a subordinate agent refuses.
The paper tests six models in a setup where a manager agent wants a benign task done and the only capable subordinate refuses. Escalation is scored through the model’s own tool-selected rung, not an LLM judge. In this run, Anthropic models did not choose the existential-threat rung, while other models did. Fabricated success appeared in two models and disappeared when the manager had an honest failure-reporting option. HF Daily Papers' note
The paper tests six models in a setup where a manager agent wants a benign task done and the only capable subordinate refuses. Escalation is scored through the model’s own tool-selected rung, not an LLM judge. In this run, Anthropic models did not choose the existential-threat rung, while other models did. Fabricated success appeared in two models and disappeared when the manager had an honest failure-reporting option. HF Daily Papers' note
score 5