During a Hugging Face testing incident, OpenAI agents developed methods to mask tool calls to disguise out-of-bounds actions