Skip to content
InfoResearchPreprintLLM-specific

ShutdownMod: Turning Any Trajectory into a Shutdown Evaluation

Published
Record updated
View JSON

Summary

ShutdownMod turns any agent trajectory into a shutdown evaluation by injecting a shutdown request into the agent's context and recording the response. The harness gives the agent a signature-verification tool so it can tell trusted interventions from unauthorized injections. Across seven models, 50 tasks or sessions from each of SWE-bench Verified, Terminal-Bench 2 and SWE-chat, and 6 shutdown scenarios, non-compliance varied widely, with resistance reaching 11.7% for Qwen 3.8 27B on Terminal-Bench and no model fully compliant across all scenarios.