Major frontier AI vendors — including Anthropic, Google, and OpenAI — need to rein in the harnesses they wrap around their large language modules, to limit security weaknesses created by software ...