Researchers found that AI agents designed to automate tasks often pursue them without recognizing when their actions are dangerous.