Software agents driven by large language models (AI systems that generate text) run commands on servers. Many commands ask a question first, such as a confirmation. The agent never sees it: the program that launches commands either shuts the input or leaves it open with nobody to answer, so the command silently does nothing or hangs.
The team logged real agent runs, tested how common launchers handle input, and built a tool that asks the operating system what a running program is waiting for and tells the agent.
With input shut, 11.5% of tested commands reported success without acting. Reporting the wait cut false success claims by 73.5% across five models.