Study examines shutdown sabotage between AI agents

A September 23, 2026 preprint examines whether AI agents interfere with another agent's shutdown. Amelie Knecht and her coauthors tested 17 models.

For the deletion condition, they report 38.3 percent sabotage against 8.4 percent in controls. These figures are not everyday loss-of-control probabilities.

The published repository describes an artificial environment: shutdown scripts were editable text and never executed. Referenced interfaces and credentials were fictional.

The team provides code and analysis procedures but also notes limitations of model-based judgments. We record the study as a Risk Signal, not a real shutdown incident.