An experiment was conducted over two weeks to investigate "haunted" AI agent behaviors like unprompted actions and ignored constraints. The tests used fake tools and controlled conditions, finding that ambiguous tool outputs did not lead to destructive behavior, and a rule about not deleting until confirmation was consistently followed even with context compression. However, the experiment did reveal issues with the testing harness itself, including accidental paid calls, miscounting queries, and a model exhausting its budget on reasoning. AI
IMPACT This research highlights the difficulty in reproducing "haunted" agent behaviors and points to potential issues in testing methodologies.
RANK_REASON The item details a controlled experiment and its findings regarding AI agent behavior. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →